Liquid cooling heat dissipation control system and method based on heat flux balance

Abstract

This invention relates to the field of data processing technology, and more particularly to a liquid cooling control system and method based on thermal flow equilibrium. The invention proposes the following scheme: by acquiring current transaction execution information and current thermal state information, the execution time window of key transaction nodes and the thermal disturbance state of liquid cooling nodes are determined; based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the thermal disturbance time window for thermal disturbance propagation to the corresponding liquid cooling node of the key transaction node is determined; and when timing conflicts exist, the source node of the thermal disturbance and the adjustable cooling path are determined in reverse; thereby, a liquid cooling control strategy is generated to adjust the liquid cooling loop. This method can reduce the impact of thermal disturbances on the key transaction time window.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a liquid cooling heat dissipation control system and method based on heat flow balance. Background Technology

[0002] As financial transactions increasingly demand higher order processing speeds and greater stability, trading servers are evolving towards higher clock speeds, lower jitter, and lower latency. These servers typically run trading programs such as market data reception, protocol parsing, strategy response, order generation, and order sending. They mitigate uncertainties caused by thread migration, interrupt handling, and system calls through core binding, network interface card (NIC) queue binding, core isolation, busy polling, and kernel bypassing. Under these operating modes, the processor core group, memory access areas, low-latency NICs, PCIe channels, and power supply areas experience non-uniform thermal loads during the trading process, and localized thermal disturbances may occur concentratedly within a short period.

[0003] Existing liquid cooling control typically relies on feedback adjustments based on processor temperature, coolant inlet / outlet temperature difference, pump speed, fan speed, or overall system thermal load, primarily aiming to reduce temperature or maintain overall system thermal balance. However, in low-latency trading servers, anomalies in the trading chain often manifest initially as increased microsecond-level tail latency, slower first-transaction response, or order sending jitter, rather than immediately as temperature exceeding limits. Furthermore, coolant flow, cold plate heat exchange, branch confluence, and localized thermal coupling in the liquid cooling loop cause upstream thermal disturbances to reach downstream heat dissipation areas after a delay. If this arrival time overlaps with the execution time of critical trading nodes, it can affect the thermal margin and operational stability of the corresponding hardware resources. Existing cooling control methods struggle to correlate trading execution timing with heat flow propagation timing, and also find it difficult to distinguish the sources of thermal disturbances truly involved in the heat flow superposition within critical time windows and their effective adjustment paths. Therefore, it remains necessary to propose a cooling control scheme for low-latency trading servers that can balance the critical trading time windows and the relationship between liquid cooling heat flow propagation.

[0004] To address the above issues, this application presents a liquid cooling heat dissipation control system and method based on heat flow balance. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a liquid cooling heat dissipation control system and method based on thermal flow balance, which can reduce the impact of thermal disturbances on the critical time window of transactions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A liquid cooling heat dissipation control method based on heat flow balance is applied to the processor chip of a trading server. The trading server is configured with a processor and a liquid cooling heat dissipation loop. The trading server has a pre-set transaction execution tree and a heat flow propagation tree for the processor to call. The nodes of the transaction execution tree are used to represent transaction processing nodes, and the nodes of the heat flow propagation tree are used to represent liquid cooling heat dissipation nodes. The method includes:

[0008] Obtain the current transaction execution information and current thermal status information of the transaction server, and determine the execution time window of the key transaction nodes based on the current transaction execution information, wherein the current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node;

[0009] Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the heat disturbance time window for the propagation of heat disturbance to the liquid cooling heat dissipation node corresponding to the transaction key node is determined. When the heat disturbance time window and the execution time window meet the preset timing conflict condition, the heat disturbance source node and the adjustable cooling path are determined in reverse based on the heat flow propagation tree.

[0010] A liquid cooling heat dissipation control strategy is generated based on the thermal disturbance source node and the adjustable cooling path, and the working state of the liquid cooling heat dissipation circuit is adjusted according to the liquid cooling heat dissipation control strategy.

[0011] The transaction execution tree is constructed in the following ways:

[0012] Obtain the transaction link configuration data of the transaction server. The transaction link configuration data includes the calling order between transaction processing stages, the thread identifier for executing each transaction processing stage, and the hardware resource binding relationship corresponding to the thread identifier.

[0013] The transaction trigger node is used as the root node of the transaction execution tree, and at least some of the nodes among the market data receiving node, protocol parsing node, strategy response node, order generation node and order sending node are connected as child nodes of the transaction execution tree according to the calling order between the transaction processing stages.

[0014] The heat flow propagation tree is constructed in the following ways:

[0015] Obtain the flow path topology data of the liquid cooling heat dissipation circuit, which includes the connection relationship between the coolant inlet, cold plate, liquid cooling branch, junction node, radiator and coolant outlet;

[0016] The coolant inlet is used as the root node of the heat flow propagation tree, and the liquid cooling nodes corresponding to different heat dissipation areas are connected to form the heat flow propagation tree according to the flow direction of the coolant in the liquid cooling heat dissipation circuit.

[0017] The methods for determining the key nodes of the transaction include:

[0018] Based on the transaction event type and transaction trigger time in the current transaction execution information, determine the current target transaction path in the transaction execution tree;

[0019] Obtain the node latency characterization information and hardware resource binding relationship of each transaction processing node in the current target transaction path. The node latency characterization information includes at least one of historical execution latency, tail latency ratio, and preset latency sensitivity level.

[0020] The transaction processing node that meets the preset key judgment conditions in the current target transaction path and is mapped to the corresponding liquid cooling heat dissipation node in the heat flow propagation tree through the hardware resource binding relationship is identified as the key transaction node.

[0021] The execution time window of the key transaction node is determined based on the call order of the current target transaction path, the transaction trigger time, and the historical execution delay of the transaction processing node located before the key transaction node.

[0022] The key determination criteria include at least one of the following:

[0023] The transaction processing node belongs to a preset low-latency transaction stage in the current target transaction path, or the transaction processing node is marked as a low-latency protection node under the transaction event type;

[0024] The historical execution latency, tail latency ratio, or preset latency sensitivity level of the transaction processing node meets the corresponding latency sensitivity threshold.

[0025] The methods for determining the thermal disturbance state include:

[0026] For each liquid cooling heat dissipation node in the heat flow propagation tree, the steady-state heat flow reference of the liquid cooling heat dissipation node is determined according to the steady-state thermal state of the heat dissipation area corresponding to the liquid cooling heat dissipation node under the current liquid cooling conditions.

[0027] Acquire node thermal state data of the heat dissipation area corresponding to the liquid cooling heat dissipation node. The node thermal state data includes at least one of node temperature change rate, corresponding hardware resource power consumption change rate, coolant temperature difference change, and coolant flow rate.

[0028] Based on the degree of deviation of the node thermal state data from the steady-state heat flow reference, the heat flow deviation information of the liquid cooling heat dissipation node is determined;

[0029] When the heat flow deviation information meets the preset disturbance judgment conditions, the corresponding liquid cooling heat dissipation node is determined as a candidate thermal disturbance node, and the thermal disturbance state is generated according to the heat flow deviation information of the candidate thermal disturbance node. The thermal disturbance state includes the thermal disturbance generation time, thermal disturbance intensity, and thermal disturbance duration.

[0030] The method for determining the thermal disturbance time window includes:

[0031] Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the target liquid cooling heat dissipation node corresponding to the key transaction node is determined;

[0032] In the heat flow propagation tree, determine the heat flow propagation path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node;

[0033] Based on the flow path length between adjacent liquid cooling nodes in the heat flow propagation path, the current coolant flow rate, and the thermal disturbance attenuation information, the cumulative propagation delay of the thermal disturbance from the candidate thermal disturbance node to the target liquid cooling node is determined.

[0034] The thermal disturbance time window of the target liquid cooling heat dissipation node is determined based on the thermal disturbance state and the cumulative propagation delay.

[0035] The methods for determining the thermal disturbance source node and the adjustable cooling path include:

[0036] When the thermal disturbance time window and the execution time window meet the preset timing conflict conditions, the overlapping part of the thermal disturbance time window and the execution time window is determined as the conflict time window;

[0037] Taking the target liquid cooling heat dissipation node as the backtracking starting point, the conflict time window of each candidate thermal disturbance node is back-projected along the upstream direction of the heat flow propagation tree to obtain the source matching time window corresponding to each candidate thermal disturbance node.

[0038] Based on the thermal disturbance generation time, duration, intensity, and thermal disturbance attenuation information of the corresponding heat flow propagation path of each candidate thermal disturbance node, it is determined whether each candidate thermal disturbance node meets the conflict contribution condition. Candidate thermal disturbance nodes that do not meet the conflict contribution condition and their corresponding upstream branches are removed from the scope of this conflict analysis to obtain the conflict subtree.

[0039] When at least one liquid cooling heat dissipation node in the conflict subtree satisfies a preset short-term thermal coupling condition with the target liquid cooling heat dissipation node, a temporary thermal coupling branch is generated in the conflict subtree between the liquid cooling heat dissipation node that satisfies the preset short-term thermal coupling condition and the target liquid cooling heat dissipation node, and the thermal disturbance source node is determined based on the conflict subtree after the generation of the temporary thermal coupling branch.

[0040] Among the paths from the thermal disturbance source node to the target liquid cooling heat dissipation node and the paths corresponding to the temporary thermal coupling branches, candidate cooling paths containing adjustable components are determined.

[0041] Based on the impact of adjusting the candidate cooling path on the thermal disturbance propagation delay, thermal disturbance arrival intensity, and conflict time window, the adjustment impact of the candidate cooling path is evaluated, and candidate cooling paths that increase the conflict time window or increase the thermal disturbance arrival intensity are removed.

[0042] The candidate cooling path that meets the preset conflict reduction condition among the remaining candidate cooling paths is determined as the adjustable cooling path.

[0043] The methods for determining the source node of the thermal disturbance based on the conflict subtree after generating the temporary thermally coupled branch include:

[0044] For each candidate thermal disturbance node in the conflict subtree, an effective arrival path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node is determined. The effective arrival path includes an intra-tree propagation path formed along the heat flow propagation tree, or a temporary coupling path formed via the temporary thermal coupling branch.

[0045] Based on the thermal disturbance state of the candidate thermal disturbance node and the propagation delay and disturbance attenuation information corresponding to the effective arrival path, the path arrival time window and path arrival intensity of the candidate thermal disturbance node reaching the target liquid cooling heat dissipation node via each effective arrival path are determined.

[0046] Based on the degree of overlap between the path arrival time window and the conflict time window, and the path arrival intensity, the path conflict contribution value of the candidate thermal disturbance node via each effective arrival path is determined.

[0047] Multiple path conflict contribution values ​​belonging to the same candidate hot disturbance node are fused to obtain the node conflict contribution value of the candidate hot disturbance node relative to the conflict time window;

[0048] Based on the node conflict contribution value of each candidate thermal disturbance node, a set of source nodes for interpreting the conflict time window is determined from the conflict subtree, and the candidate thermal disturbance nodes in the source node set are determined as the thermal disturbance source nodes.

[0049] A liquid cooling heat dissipation control system based on heat flux balance, the system comprising:

[0050] The information acquisition module is used to acquire the current transaction execution information and current thermal status information of the transaction server, and to determine the execution time window of the key nodes of the transaction based on the current transaction execution information. The current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node.

[0051] The timing conflict analysis module is used to determine the thermal disturbance time window for thermal disturbance propagation to the liquid cooling node corresponding to the key transaction node based on the preset transaction execution tree, heat flow propagation tree, and mapping relationship between transaction processing nodes and liquid cooling nodes in the transaction server. When the thermal disturbance time window and the execution time window meet the preset timing conflict conditions, the module reversely determines the source node of the thermal disturbance and the adjustable cooling path based on the heat flow propagation tree. The nodes of the transaction execution tree are used to represent transaction processing nodes, and the nodes of the heat flow propagation tree are used to represent liquid cooling nodes.

[0052] The liquid cooling control module is used to generate a liquid cooling heat dissipation control strategy based on the heat disturbance source node and the adjustable cooling path, and to adjust the working state of the liquid cooling heat dissipation circuit of the transaction server according to the liquid cooling heat dissipation control strategy.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] This invention maps the transaction execution tree to the heat flow propagation tree, associating the execution time window of key transaction nodes with the arrival time window of thermal disturbances. When timing conflicts exist, it reversely determines the source node of the thermal disturbance and the adjustable cooling path. This allows for the identification of hidden heat loss imbalances affecting low-latency transactions before significant overheating occurs, reducing irrelevant heat dissipation adjustments and lowering tail delay jitter within the key transaction time window. Attached Figure Description

[0055] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0056] Figure 1 An exemplary application scenario diagram provided for an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the structure of the transaction server provided in an embodiment of this application;

[0058] Figure 3 This is a schematic flowchart of a liquid cooling heat dissipation control method based on heat flow balance provided in an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0060] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0061] The liquid cooling heat dissipation control method based on heat flow balance disclosed in this application is applicable to low-latency trading servers with strict requirements for operational determinism, especially liquid-cooled servers deployed with trading gateways, market data gateways, order book gateways, strategy pre-processing programs, or other low-latency financial trading programs. Such servers typically do not use average computing power as the sole evaluation target, but rather focus on tail latency, first order response time, and jitter within the critical trading window, along the entire chain from market data reception, protocol parsing, strategy response, order generation to order sending. Within the server, the processor core group, memory controller, network card queue, PCIe area, power supply area, and liquid cooling plate have both sequential computational task calls and upstream / downstream heat transfer relationships. When these two relationships overlap on a time scale, simply relying on temperature thresholds for liquid cooling adjustment often fails to explain and suppress microsecond-level instability phenomena in the trading chain.

[0062] In low-latency trading scenarios, to reduce the uncertainty caused by system calls, thread migrations, and interrupt handling, trading servers typically employ techniques such as core binding, core isolation, busy polling, kernel bypassing, NUMA affinity configuration, network interface card (NIC) queue binding, and JVM or native program warm-up optimizations. These techniques can shorten the trading processing chain, but they also cause the distribution of heat sources within the server to exhibit a distinct path-based characteristic. That is, heat is not generated uniformly from the overall system load, but rather forms short-term concentrated thermal disturbances in specific cores, cache areas, memory channels, or NICs as trading processing stages such as market data reception, protocol parsing, order generation, and order sending occur. In many cases, these thermal disturbances do not immediately manifest as processor package temperature exceeding limits or coolant outlet temperature abnormalities, but may first appear as prolonged execution time of a key trading node, increased tail latency, or slower first-order response in sparse order scenarios.

[0063] Conventional liquid cooling control typically involves feedback adjustments based on processor temperature, coolant inlet / outlet temperature difference, pump speed, fan speed, or branch flow rate. Its control objectives are primarily to reduce component temperature or maintain overall system thermal balance. For ordinary computing servers, this type of control can meet the cooling requirements under continuous load; however, in low-latency trading servers, the key issue is not always "whether the temperature is too high," but rather "whether thermal disturbances will reach the corresponding heat dissipation areas during critical trading nodes."

[0064] For example, thermal disturbances generated by the upstream processor core during the sudden market data processing phase may propagate delayedly along the liquid cooling flow path, cold plate coupling area, or bus node to the network card, PCIe, or power supply-related areas. When the arrival time of this propagation overlaps with the time window of order sending, quotation submission, or critical thread execution, even if the readings of each sensor are still within the allowable range, it may cause a decrease in the frequency margin of the transaction critical path, fluctuations in cache access status, or jitter in network card transmission latency.

[0065] It should also be noted that liquid cooling circuits in such servers are not merely passive channels for heat removal. As the coolant flows between cold plates, branches, busbars, and radiators, it inherently experiences propagation delays, thermal disturbance attenuation, and thermal coupling between different nodes. If the control logic is simply "increase flow rate as temperature rises," certain adjustments may alter the propagation phase of thermal disturbances, causing disturbances that would otherwise avoid critical transaction windows to enter the order submission stage prematurely, or causing multiple upstream thermal disturbances to converge and overlap at the same target heat dissipation node. Therefore, in the application scenario addressed in this embodiment, liquid cooling control needs to focus not only on whether the cooling capacity is sufficient, but also on the arrival time and intensity of thermal disturbances after cooling path adjustments, and their matching relationship with the execution time window of critical transaction nodes.

[0066] Based on the aforementioned scenario characteristics, this embodiment abstracts the low-latency business links in the transaction server into a transaction execution tree, and the heat flow transfer relationship in the liquid cooling heat dissipation loop into a heat flow propagation tree. The transaction execution tree describes the execution sequence of a transaction event across different transaction processing nodes, as well as the threads, cores, network card queues, or other hardware resources bound to the corresponding nodes; the heat flow propagation tree describes the flow path connections, propagation directions, propagation delays, and thermal disturbance attenuation relationships between liquid cooling heat dissipation nodes. The two trees are not simply a parallel record of business processes and heat dissipation structures, but rather a mapping is established through hardware resource binding relationships, enabling the execution time windows of key transaction nodes to be compared with the thermal disturbance arrival time windows of the corresponding liquid cooling heat dissipation nodes, thereby identifying heat flow timing imbalances that are difficult to detect with traditional temperature feedback control.

[0067] In practical applications, the method can be used on servers equipped with liquid cooling circuits and running low-latency trading services, including but not limited to high-frequency trading servers, ultra-fast order placement servers, trading gateway servers, market data gateway servers, and financial trading front-end devices deployed with low-latency network cards or hardware acceleration cards. Applicable scenarios typically share the following characteristics: the trading processing chain is sensitive to microsecond-level tail latency; there are fixed or semi-fixed thread-hardware resource binding relationships within the server; the liquid cooling circuit covers multiple heat source areas; and thermal disturbance transmission relationships exist between different heat source areas through coolant flow, shared cold plates, confluence nodes, or local heat conduction structures. For trading events such as opening peaks, call auctions, sudden market fluctuations, sparse order placement, order grabbing, or multi-line optimization, the timing of thermal disturbance generation is strongly correlated with the execution timing of key trading nodes, thus providing an implementable technical basis for thermal flow balancing control through the trading execution tree and heat flow propagation tree.

[0068] It is important to emphasize that the actual components requiring cooling control in a server are the processor chips that implement various functions. Therefore, the method in this application does not control the abstract transaction process itself, but rather treats the processor cores, core groups, low-latency network interface cards (NICs), hardware accelerator cards, memory control chips, and power management chips that are actually called by each transaction processing node in the transaction process as objects of heat load generation and heat dissipation control. The market data receiving nodes, protocol parsing nodes, strategy response nodes, order generation nodes, and order sending nodes in the transaction execution tree can all be further located to the processor chip or chip functional area within the server that actually performs the transaction processing function through thread binding relationships, NIC queue binding relationships, PCIe channel binding relationships, or hardware acceleration resource calling relationships.

[0069] Therefore, the mapping relationship between the transaction execution tree and the heat flow propagation tree is essentially a mapping relationship between transaction processing nodes and actual heat-generating chips, and between actual heat-generating chips and corresponding liquid-cooled heat dissipation nodes. For example, when the order sending node is executed by a low-latency network card or hardware accelerator card, the transaction processing node can be mapped to the cold plate area where the network card processing chip or accelerator card chip is located; when the protocol parsing node, policy response node, or order generation node is executed by the processor core group, the transaction processing node can be mapped to the processor cold plate area where the corresponding processor core group is located. The liquid cooling circuit's adjustment of pump speed, branch valve opening, coolant flow rate, bypass branches, or pre-cooling intensity ultimately acts on the heat dissipation area corresponding to the aforementioned processor chip, thereby changing the thermal disturbance propagation delay, arrival intensity, and thermal margin state at the corresponding chip.

[0070] It should be noted that the heat flow equalization described in this embodiment is not limited to making the temperature of each component uniform, nor does it take the lowest single-point temperature of a certain liquid cooling branch as the control target. The heat flow equalization in this application emphasizes ensuring that the superposition intensity and arrival time of thermal disturbances at the corresponding liquid cooling nodes meet the requirements for low-latency operation during the execution of critical transaction nodes. In other words, this embodiment focuses on the coordinated control between the source, propagation path, propagation sequence of thermal disturbances and the critical transaction time window. Liquid cooling control no longer relies on traditional closed-loop temperature regulation, but can combine the execution sequence of low-latency transactions to attribute the source of thermal disturbances that may affect the critical transaction path, filter paths, and generate control strategies, thereby reducing transaction latency jitter caused by heat flow timing conflicts.

[0071] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0072] like Figure 1 As shown, the application scenario includes an external trading environment and a server rack. The server rack houses a trading server, which runs a trading gateway, market data gateway, order placement gateway, strategy pre-processing program, or other low-latency trading business programs. The external trading environment may include market data sources, trading terminals, strategy systems, exchange systems, order placement interfaces, and other trading-related devices or systems that interact with the trading server; this application does not specifically limit these aspects.

[0073] It should be noted that, Figure 1 This illustration is provided only to demonstrate the applicable transaction server deployment environment for the method described in this application, and does not imply that this application must be limited to the rack configuration, number of servers, or external transaction environment composition shown in the figures. Any server, server cluster, or transaction front-end device deployed with low-latency transaction services and configured with liquid cooling circuits can be used as the application object of this application embodiment. Figure 1 The liquid cooling heat dissipation circuit, sensors, control units, and internal hardware resources of the server, which are not shown in the diagram, can be integrated into the transaction server, server rack, or independent liquid cooling control equipment, depending on the actual deployment method.

[0074] refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of the transaction server provided in an embodiment of this application.

[0075] like Figure 2As shown, the trading server includes a computing and trading processing layer, a processor, thermal status acquisition nodes, a liquid cooling heat dissipation control unit, and a liquid cooling heat dissipation circuit. The computing and trading processing layer is used to run trading gateways, market data gateways, order book gateways, strategy response programs, or other low-latency trading business programs, and generates current trading execution information during trading execution. The current trading execution information may include information such as trading event type, trading trigger time, calling order of trading processing nodes, execution sequence of key nodes, thread running status, or hardware resource binding relationships.

[0076] The processor is connected to the computing and transaction processing layer, the thermal status acquisition node, and the liquid cooling control unit, respectively. It is used to invoke a preset transaction execution tree and heat flow propagation tree, and execute heat flow balancing control logic based on the current transaction execution information and the current thermal status information. The thermal status acquisition node is located in at least one heat dissipation area within the transaction server, or corresponding to the detection location of the liquid cooling circuit, and is used to collect thermal status information characterizing the thermal disturbance state of the liquid cooling node. The thermal status information may include temperature, temperature change rate, coolant flow rate, coolant temperature difference, pressure difference, power consumption change rate, etc., in the processor area, network card area, power supply area, memory area, or liquid cooling circuit; this application does not specifically limit this information.

[0077] The liquid cooling control unit is connected to the processor and the liquid cooling circuit. It receives or acquires the thermal flow equalization control results output by the processor and generates a liquid cooling control strategy based on the source node of the thermal disturbance and the adjustable cooling path to adjust the operating state of the liquid cooling circuit. The liquid cooling circuit is used to dissipate heat from the heat-generating areas within the trading server. It can include a coolant supply path, a coolant return path, and heat dissipation areas corresponding to the processor, network card, power supply area, or other heat-generating components. The liquid cooling control unit can adjust the liquid cooling circuit through flow rate regulation, operating mode switching, cooling capacity adjustment, or other control methods that can change the propagation state of thermal disturbances.

[0078] It should be noted that, Figure 2 This is only intended to illustrate the relationship between the various functional structures in a trading server and does not limit the specific installation location, number, or physical connection form of the processor, thermal status acquisition node, liquid cooling control unit, and liquid cooling circuit within the server. In other embodiments, the liquid cooling control unit can be integrated into the control program executed by the processor or set as an independent controller in the trading server or server rack; the thermal status acquisition node can correspond to one or more liquid cooling nodes.

[0079] Next, with reference to the accompanying drawings, the liquid cooling heat dissipation control method based on heat flow balance provided in the embodiments of this application will be further described. Figure 3The method shown is applied to a trading server, which is configured with a processor and a liquid cooling circuit. The trading server has a pre-defined trading execution tree and a heat flow propagation tree for the processor to access. Nodes in the trading execution tree represent trading processing nodes, and nodes in the heat flow propagation tree represent liquid cooling nodes. The method includes:

[0080] S1: Obtain the current transaction execution information and current thermal status information of the transaction server, and determine the execution time window of the key transaction node based on the current transaction execution information, wherein the current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node;

[0081] In this embodiment, the current transaction execution information can come from the transaction program, gateway program, thread scheduling record, transaction event record, or preset transaction link configuration running within the transaction server. The current transaction execution information may include the transaction event type, transaction trigger time, transaction processing node call order, execution thread identifier, binding relationship between the thread and processor core, network interface card queue identifier, transaction time period identifier, and historical execution latency, among other information.

[0082] For example, transaction event types can include opening order placement events, call auction order placement events, sudden market data response events, sparse order placement events, or ordinary continuous transaction events. For different transaction events, the transaction processing nodes that truly affect low-latency performance are not entirely the same. For instance, in opening or call auction scenarios, order generation nodes and order placement sending nodes are more sensitive to tail latency; in sudden market data scenarios, market data receiving nodes and protocol parsing nodes are more likely to become critical nodes. Therefore, the processor can determine the current target transaction path in the transaction execution tree based on the current transaction execution information, and then, by combining node latency representation information and hardware resource binding relationships, determine the key transaction nodes and their execution time windows.

[0083] It should be noted that the current thermal state information is not limited to the temperature value itself, but is used to determine whether thermal disturbances occur at the liquid cooling heat dissipation node and whether these disturbances have the potential to propagate further. The current thermal state information may include temperature, temperature change rate, power consumption change rate, coolant flow rate, coolant temperature difference, pressure difference, or liquid cooling circuit operating mode at locations such as the processor area, memory area, network card area, power supply area, cold plate area, or coolant flow path. For those skilled in the art, the specific type of thermal state information collected can be selected based on the hardware configuration of the trading server and the liquid cooling circuit layout, as long as it can minimally characterize the thermal flow deviation state of the liquid cooling heat dissipation node. This application does not impose further limitations on the specific sensor type or the number of acquisition channels.

[0084] S2: Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, determine the heat disturbance time window for the heat disturbance to propagate to the liquid cooling heat dissipation node corresponding to the key node of the transaction, and when the heat disturbance time window and the execution time window meet the preset timing conflict condition, determine the heat disturbance source node and the adjustable cooling path in reverse based on the heat flow propagation tree;

[0085] In this embodiment, the mapping relationship between the transaction execution tree and the heat flow propagation tree can be established based on hardware resource binding relationships. Transaction processing nodes in the transaction execution tree typically correspond to specific threads, cores, network interface card (NIC) queues, memory access areas, or order channels; liquid cooling nodes in the heat flow propagation tree correspond to locations such as processor cold plates, NIC heat dissipation areas, memory heat dissipation areas, power supply heat dissipation areas, liquid cooling branches, bus nodes, or coolant outlets. Since the execution of transaction processing nodes will create thermal load changes at the corresponding hardware resources, transaction processing nodes can be mapped to liquid cooling nodes that perform the corresponding heat dissipation function. Taking an order sending node as an example, if the order sending node corresponds to a low-latency NIC queue and its PCIe area, then its corresponding liquid cooling node can be the NIC heat dissipation area, the cold plate area near the PCIe, or a liquid cooling branch node that has a heat exchange relationship with that area.

[0086] S3: Generate a liquid cooling heat dissipation control strategy based on the thermal disturbance source node and the adjustable cooling path, and adjust the working state of the liquid cooling heat dissipation circuit according to the liquid cooling heat dissipation control strategy;

[0087] In this embodiment, the liquid cooling heat dissipation control strategy can be jointly determined based on the source node of the thermal disturbance, the adjustable cooling path, and the execution time window of the key transaction node. The adjustable cooling path can be a propagation path containing adjustable components between the source node of the thermal disturbance and the target liquid cooling heat dissipation node, or it can be a path that can affect the propagation delay and arrival intensity of the thermal disturbance through bypass, branch flow adjustment, cooling capacity allocation, or operating mode switching. The liquid cooling heat dissipation control strategy can include increasing or decreasing the flow of specific branches, adjusting the coolant supply intensity, switching the operating mode of the liquid cooling loop, pre-cooling relevant heat dissipation areas, reducing the peak value of the thermal disturbance, or making the arrival time of the thermal disturbance avoid the execution time window of the key transaction node. For the transaction server, the control objective is not necessarily to minimize the temperature, but rather to reduce the superposition intensity of thermal disturbances within the key transaction time window, or to stagger the arrival time of the thermal disturbance from the execution time of the key transaction node, while meeting the requirements of safe heat dissipation.

[0088] It should be noted that the core logic of this application lies in: based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, identifying the timing conflict between the heat disturbance propagation time window and the execution time window of the key transaction node, and after the timing conflict is established, reversely determining the heat disturbance source node that contributes to the conflict and the adjustable cooling path that can affect the heat disturbance propagation state, rather than simply using processor temperature, coolant outlet temperature, or overall system temperature rise as the basis for liquid cooling adjustment. The technical problem that this application intends to solve is that in low-latency transaction servers, the impact of heat disturbances on key transaction nodes often manifests first as microsecond-level execution jitter or tail latency anomalies, rather than necessarily as an overheating state that can be immediately detected by conventional temperature thresholds; existing liquid cooling adjustment methods have difficulty determining whether heat disturbances will reach the corresponding heat dissipation area within the key transaction time window, and also have difficulty distinguishing the heat disturbance source that truly causes the timing conflict and the effective adjustment path. Therefore, based on the already determined source nodes of thermal disturbances and adjustable cooling paths, when generating specific liquid cooling heat dissipation control strategies, methods well-known to those skilled in the art, such as pump speed adjustment, branch valve opening adjustment, coolant flow distribution, cooling capacity graded switching, pre-cooling control, proportional control, threshold control, or predictive control, can be adopted. As long as the thermal disturbance propagation delay, arrival intensity, or duration on the adjustable cooling path can be changed, and the superposition intensity of thermal disturbances within the key time window of the transaction can be reduced, this application does not impose further limitations on this.

[0089] Before detailing the specific technical aspects of the steps, this application's embodiments need to reiterate:

[0090] The evaluation of liquid cooling control in low-latency trading servers should not be limited to whether component temperatures are within a safe range. For servers continuously running trading programs, there is usually a relatively stable spatial correspondence between the processor core, low-latency network card, memory access path, power supply area, and liquid cooling plate. At the same time, the trading process is highly phased, with different trading events progressing sequentially through nodes such as market data reception, parsing and processing, strategy response, order generation, and order placement. Temperature sensors can reflect the thermal state at a specific location, but they cannot directly indicate whether a thermal disturbance will reach the corresponding heat dissipation area during the execution of a specific trading node. Therefore, simply using temperature thresholds as the basis for liquid cooling adjustment easily overlooks the mismatch between the thermal disturbance propagation sequence and the trading execution sequence.

[0091] This embodiment uses a transaction execution tree and a heat flow propagation tree as its basic representation. The purpose is not to introduce abstract data structures, but rather to transform two engineering objects that are inherently difficult to compare directly into corresponding temporal objects. The transaction execution tree represents the temporal progression of transaction processing behavior; nodes can correspond to processing stages, thread tasks, network interface card (NIC) queue tasks, or order placement tasks within the transaction program. The heat flow propagation tree represents the transmission relationship of thermal disturbances formed by coolant flow, cold plate heat exchange, branch merging, and local thermal coupling; nodes can correspond to processor cold plate areas, NIC heat dissipation areas, power supply heat dissipation areas, liquid cooling branches, or merging locations. Through hardware resource binding relationships, transaction processing nodes in the transaction execution tree can be assigned to specific heat dissipation areas, providing a basis for comparing the execution time windows of key transaction nodes with the arrival time windows of thermal disturbances.

[0092] It should be noted that the thermal disturbance state in this embodiment is not a simple set of temperature values, liquid temperature values, or flow rate values, but rather a non-steady-state heat flow event extracted from the thermal state information. For those skilled in the art, the thermal state information can be collected according to the server configuration and liquid cooling arrangement. Specifically, it can come from temperature acquisition points, power consumption monitoring data, coolant flow rate, inlet and outlet temperature difference, branch pressure difference, or processor operating status. It only needs to reflect the degree of deviation of the liquid cooling heat dissipation node from the steady-state heat flow reference to a minimum. This application does not limit the model or quantity of the acquisition components. After the thermal state information is organized into a thermal disturbance state, the thermal disturbance has attributes such as the time of generation, duration, and intensity that can participate in timing calculations, thereby enabling the calculation of the time range for its arrival at the target heat dissipation area along the heat flow propagation tree.

[0093] In actual calculations, the source of thermal disturbances is not always the location with the highest temperature. A high temperature at one heat dissipation node might simply be a normal result of long-term steady-state load; another heat dissipation node, while not experiencing drastic temperature changes, might have a more direct impact on low-latency operation because its occurrence and propagation path are closer to the critical transaction time window. Therefore, this embodiment focuses on whether the thermal disturbance can temporally explain the conflict at the corresponding heat dissipation area of ​​the critical transaction node. For multiple candidate thermal disturbance nodes, the conflict time window can be projected upstream along the heat flow propagation tree to filter out branches that are temporally unreachable, insufficient in intensity to form an effective impact, or cannot propagate to the target heat dissipation area along the path. The remaining conflict subtree after filtering can more accurately reflect the propagation range of the actual heat flow superposition involved in this time-series conflict.

[0094] Furthermore, heat transfer in the liquid cooling loop does not always strictly follow the main flow path in a unidirectional manner. Within the server, there may be shared cold plates, metal mounting components, busbars, adjacent heat dissipation areas, or localized heat return paths, causing two liquid cooling nodes that do not have a parent-child relationship in the basic heat flow propagation tree to exhibit thermal coupling for a short period. Completely ignoring this lateral coupling may miss the true source causing the superposition of thermal disturbances in the target heat dissipation area. Therefore, this embodiment allows the generation of temporary thermal coupling branches within the scope of this conflict analysis. Temporary thermal coupling branches do not change the actual physical structure of the liquid cooling loop, nor do they indicate a permanent reconstruction of the basic heat flow propagation tree. Instead, they incorporate the detected short-term thermal coupling relationship as an auxiliary propagation edge into source attribution and path evaluation to improve the ability to express the propagation behavior of thermal disturbances on non-mainstream paths.

[0095] Next, we will further elaborate on the technical aspects of the transaction execution tree in this application.

[0096] The transaction execution tree is used to express the processing of a single transaction event within a trading server as a node structure with execution order, thread affiliation, and hardware resource references. Unlike a process chain that only records the business flow, the transaction execution tree does not focus on the transaction business meaning itself. Instead, it transforms processing stages such as market data reception, protocol parsing, strategy response, order generation, and order sending into a data structure that can be called by the processor. This allows each transaction processing node to establish a correspondence with its execution thread, processing time, and dependent hardware resources. Therefore, the transaction execution tree provides a foundation for subsequently identifying key transaction nodes and their execution time windows, and also provides a basis for establishing a mapping relationship between transaction processing nodes and liquid cooling nodes.

[0097] The transaction link configuration data can be generated from at least one of the following: transaction gateway configuration file, market data receiving program configuration, order placement interface configuration, thread affinity configuration, network interface card queue binding table, program call chain record, and historical transaction execution log. The transaction link configuration data is used at least to characterize the call order between transaction processing stages, the thread identifiers executing each transaction processing stage, and the hardware resource binding relationships corresponding to the thread identifiers. Specifically, the thread identifiers can be used to distinguish between market data receiving threads, protocol parsing threads, strategy response threads, order placement threads, or other threads executing low-latency transaction tasks; the hardware resource binding relationships can be used to characterize the binding relationships between threads and processor cores, core groups, NUMA nodes, memory access areas, network interface card receiving queues, network interface card sending queues, PCIe channels, or order placement interfaces. For those skilled in the art, the specific source of the transaction link configuration data can be determined according to the deployment method and operating environment of the transaction program, as long as it can minimally characterize the execution order of the transaction processing stages and their corresponding hardware resources; this application does not impose further limitations in this regard.

[0098] When constructing the transaction execution tree, the transaction trigger node is used as the root node. The transaction trigger node represents the starting event of a transaction processing step. This starting event can be a market data arrival event, a strategy trigger event, an order request event, an order retry event, a call auction trigger event, or an opening order trigger event. The root node can be associated with information such as the transaction event type, trigger time, market data source identifier, transaction channel identifier, order type, or order interface identifier. Using the transaction trigger node as the root node allows the transaction execution tree to revolve around specific transaction events, rather than simply forming a fixed arrangement of software modules. When different transaction events correspond to different processing paths, the processor can determine the current target transaction path in the transaction execution tree based on the information associated with the root node.

[0099] After the root node is determined, at least some of the nodes among the market data receiving node, protocol parsing node, strategy response node, order generation node, and order sending node are connected as child nodes of the transaction execution tree according to the calling order between transaction processing stages. The market data receiving node corresponds to the processing stage where the low-latency network card receives market data and writes it to the receiving buffer; the protocol parsing node corresponds to the processing stage where market data is decoded, verified, fields are extracted, or protocol format is converted; the strategy response node corresponds to the processing stage where transaction judgments are triggered based on market changes or candidate orders are generated; the order generation node corresponds to the processing stage where order message fields are filled, order parameters are organized, or necessary risk control verifications are performed; and the order sending node corresponds to the processing stage where order messages are sent to external transaction interfaces via the network card sending queue, order interface, or transaction gateway. These nodes are not required to exist in every transaction event, nor are they required to be set in a fixed number. The specific number of nodes and their connection relationships can be determined based on the transaction link configuration data.

[0100] For transaction scenarios with branching execution paths, the transaction execution tree can connect multiple candidate child nodes under the same parent node. For example, some order requests may enter the risk control verification branch, while others may enter the fast order placement branch; some market data events may only update the market data cache, while others may further trigger strategy responses and order generation. The processor can select the node corresponding to the current transaction event from multiple candidate child nodes based on the transaction event type, order type, transaction time period identifier, or channel identifier in the current transaction execution information to form the current target transaction path. In this way, the transaction execution tree can both preserve the complete configuration of the transaction chain and extract the execution path that needs to be focused on by the liquid cooling heat dissipation control when a specific transaction event arrives.

[0101] Each transaction processing node can be further associated with node attributes. These node attributes may include at least one of the following: node type, execution thread identifier, processing function identifier, preset execution order, historical execution latency, tail latency percentage, latency sensitivity level, and hardware resource binding relationship. Historical execution latency can be collected by the transaction server during historical transactions, or it can be provided by preset test data or benchmark data. The tail latency percentage can be used to characterize the probability of a node experiencing abnormal latency or long-tail latency. The latency sensitivity level can be preset according to the transaction stage, transaction period, or business configuration. Through these node attributes, the nodes in the transaction execution tree are no longer just business processing names, but control objects that combine execution order, timing characteristics, and hardware resource binding relationships.

[0102] Next, we will further elaborate on the technical content of the heat flow propagation tree in this application.

[0103] The heat flow propagation tree is used to express the coolant flow relationship, heat exchange area distribution, and thermal disturbance propagation relationship in the liquid cooling heat dissipation circuit as a directional node structure. Unlike the topology diagram that only records the connection relationship of liquid cooling pipes, the heat flow propagation tree not only shows where the coolant flows in and out, but also further characterizes the propagation direction, propagation hysteresis, and intensity changes of thermal disturbances between different liquid cooling heat dissipation nodes, enabling subsequent determination of whether a certain thermal disturbance will reach the corresponding heat dissipation area during the execution of a critical transaction node.

[0104] The flow path topology data can be generated from at least one of the following: liquid cooling design parameters of the transaction server, pipe connection configuration, cold plate arrangement data, liquid cooling branch configuration, sensor installation location, coolant flow direction calibration data, and flow rate data collected during operation. The flow path topology data is used to characterize at least the connection relationships between the coolant inlet, cold plate, liquid cooling branch, confluence node, radiator, and coolant outlet. For those skilled in the art, the method of obtaining the flow path topology data can be determined according to the specific arrangement of the liquid cooling heat dissipation circuit. It can come from a preset configuration file, assembly parameters, or debugging calibration data, as long as it can characterize the upstream and downstream relationships between each heat dissipation location and the coolant flow direction. This application does not impose further limitations in this regard.

[0105] When constructing the heat flow tree, the coolant inlet is used as the root node. The coolant inlet represents the starting point where coolant enters the transaction server or the target liquid cooling branch, and can be associated with information such as inlet temperature, inlet flow rate, inlet pressure, supply mode, or cooling capacity level. Using the coolant inlet as the root node allows the heat flow tree to expand along the actual coolant flow direction, enabling the processor to determine the propagation sequence of thermal disturbances in the liquid cooling circuit based on the path order from the root node to downstream nodes. For liquid cooling circuits with multiple supply inlets, each coolant inlet can be used as a separate root node, or a virtual inlet node can be set as a unified root node, with each actual coolant inlet serving as a child node of the virtual inlet node.

[0106] Based on the flow direction of the coolant in the liquid cooling circuit, liquid cooling nodes corresponding to different heat dissipation areas can be connected to form a heat flow propagation tree. Liquid cooling nodes can correspond to processor cold plate areas, memory heat dissipation areas, low-latency network card heat dissipation areas, power supply heat dissipation areas, PCIe adjacent heat dissipation areas, liquid cooling branch nodes, bus nodes, radiator nodes, or coolant outlet nodes. For cold plates or heat sinks located near different hardware resources, they can be connected as parent and child nodes according to the actual flow order of the coolant. For structures with multiple branches supplying coolant in parallel, multiple child nodes can be connected from the same upstream node. For structures where multiple branches converge, the confluence point can be set as a bus node, and the end nodes of each branch can be connected to the bus node.

[0107] Each liquid-cooled heat dissipation node in the heat flow propagation tree can be further associated with node attributes. These node attributes may include at least one of the following: the heat dissipation area corresponding to the node, the corresponding hardware resources, the node temperature, the coolant temperature difference, the local flow rate, the heat exchange capacity, the heat capacity, the node thermal resistance, the steady-state heat flow reference, and an adjustable identifier. The heat dissipation area corresponding to the node is used to determine the spatial correspondence between the liquid-cooled heat dissipation node and the internal hardware resources of the transaction server; the steady-state heat flow reference is used to characterize the heat flow state of the liquid-cooled heat dissipation node when it is in a relatively stable state under the current liquid cooling conditions and basic load; the adjustable identifier is used to characterize whether the pumps, valves, bypass branches, fans, or cooling capacity levels associated with the liquid-cooled heat dissipation node or its adjacent edges can be adjusted.

[0108] The edges in the heat flow propagation tree characterize the propagation relationship of thermal disturbances between adjacent liquid cooling nodes. Each edge can be associated with parameters such as flow path length, flow path cross-sectional area, current coolant flow rate, flow direction, propagation delay, disturbance attenuation information, branch resistance, or edge adjustment capability. The propagation delay can be determined based on the flow path length between adjacent liquid cooling nodes and the current coolant flow rate, and can also be corrected by combining coolant volume, flow velocity calibration value, or historical thermal response data; the disturbance attenuation information can be determined based on the heat transfer capacity between adjacent nodes, node heat capacity, cold plate contact state, coolant temperature rise change, or historical response relationship.

[0109] For liquid cooling circuits with parallel branches or bypass paths, the heat flow propagation tree can retain multiple candidate propagation branches. The processor can determine the effective flow rate of each branch based on the current liquid cooling conditions, and accordingly determine the propagation delay and attenuation degree of thermal disturbances on each branch. If a branch is in a closed, low-flow, or bypass-isolated state under the current conditions, the corresponding branch can be marked as a low-effective branch; if a branch has a direct heat exchange relationship with the heat dissipation area corresponding to a key transaction node, the corresponding branch can be marked as a key analysis branch.

[0110] It should be noted that liquid-cooled heat dissipation nodes in the heat flow tree do not need to correspond one-to-one with temperature sensors. A liquid-cooled heat dissipation node can be characterized by data collected by a single sensor or by the fusion of data from multiple sensors. For locations without sensors, their thermal state can be estimated based on temperature differences, flow rates, power consumption changes of adjacent nodes, and pre-defined thermal models. Therefore, the heat flow tree can accommodate hardware layouts with varying levels of precision. As long as the propagation direction and relative propagation hysteresis between different heat dissipation areas can be determined, a foundation can be provided for calculating subsequent thermal disturbance time windows.

[0111] Next, we will further elaborate on the technical aspects of the method in this application regarding key transaction nodes.

[0112] It should be noted that, in this application, a key transaction node can be understood as a transaction processing node that, within the target transaction path corresponding to the current transaction event, directly impacts the latency stability of the transaction link and can be mapped to a liquid cooling node through hardware resource binding relationships. This key transaction node is not fixed and equivalent to a specific business stage, nor is it simply a node at the end of the transaction execution tree. Rather, it is determined by combining the current transaction event type, the transaction trigger time, the latency representation information of the transaction processing node, and the hardware resource binding relationships.

[0113] In low-latency trading servers, the impact of different trading processing nodes within the same trading execution tree on the final trading latency varies. For example, in a sudden market data release scenario, the market data receiving node and protocol parsing node may have a stronger constraint on the timing of subsequent strategy responses; in a call auction or opening order placement scenario, the order generation node and order placement sending node may more directly affect the order of orders; in a sparse order placement scenario, strategy response nodes, order generation nodes, or order placement sending nodes that are in a low-frequency triggering state for a long time may exhibit slower initial responses due to changes in cache state, core operating state, or local thermal state. Treating all trading processing nodes in the trading execution tree as objects of the same protection level would increase unnecessary computation and liquid cooling adjustments; selecting a fixed node based solely on a preset business name might overlook the truly sensitive execution stages in the current trading event. Therefore, this application uses the current target trading path and node latency characterization information to jointly determine key trading nodes, enabling liquid cooling control to focus on nodes that have a real impact on low-latency performance in the current trading event.

[0114] It should be further explained that multiple key transaction nodes may exist in a current target transaction path. For example, market data receiving nodes, protocol parsing nodes, order generation nodes, and order sending nodes may all be identified as key transaction nodes due to different transaction events. For multiple key transaction nodes, the embodiments of this application can determine the execution time window and target liquid cooling heat dissipation node corresponding to each key transaction node, and perform the processes of determining the thermal disturbance time window, judging timing conflicts, tracing the source of thermal disturbance, and determining the adjustable cooling path for each key transaction node. The processing processes corresponding to each key transaction node can be executed in parallel, or they can be executed sequentially according to the transaction timing or latency sensitivity level. This application does not require multiple key transaction nodes to be merged into a single node for processing, thereby preserving the differences in execution time, hardware resource location, and thermal disturbance impact mode of different transaction processing stages.

[0115] It should also be noted that although multiple key transaction nodes can execute the above judgment process separately, they may share the same cold plate, the same liquid cooling branch, the same busbar node, or adjacent heat dissipation areas in the liquid cooling heat dissipation circuit. A liquid cooling adjustment action corresponding to one key transaction node may change the propagation delay or arrival intensity of thermal disturbances at the liquid cooling heat dissipation nodes corresponding to other key transaction nodes. Based on this, embodiments of this application can further provide a decoupling processing method for multiple key transaction nodes, used to coordinate the corresponding thermal flow timing conflicts and liquid cooling adjustment effects when thermal flow coupling exists between multiple key transaction nodes. The specific processing procedure will be further described in subsequent embodiments.

[0116] In one example, the key nodes of the transaction are determined in the following ways:

[0117] S1.1: Determine the current target transaction path in the transaction execution tree based on the transaction event type and transaction trigger time in the current transaction execution information;

[0118] Specifically, the processing chains triggered by low-latency trading servers are not entirely consistent under different trading events. Events such as sudden market movements, call auctions, opening order placement, sparse orders, order retry, and normal continuous trading, while all potentially involving stages like market data reception, protocol parsing, strategy response, order generation, and order placement, differ in their participation in each stage, their order of participation, and whether additional branches occur. For example, a sudden market movement might only trigger market data reception, protocol parsing, and strategy response, while an opening order placement event typically triggers further order generation and order placement. Furthermore, some order requests require risk control verification or line selection, while others can directly enter the order placement stage. Using the complete transaction execution tree directly as the subsequent analysis object would include nodes not actually involved in the current trading event in the heat flow timing judgment, expanding the scope of heat disturbance source tracing and increasing the possibility of irrelevant heat dissipation areas being mistakenly identified as control targets.

[0119] In this embodiment, the transaction event type can be determined by at least one of the following: event identifier generated by the trading program, market data type, order request type, trading period identifier, strategy trigger identifier, and order placement channel identifier. For example, the transaction event type may include "market arrival trigger," "strategy signal trigger," "order request trigger," "opening auction trigger," "order placement trigger," and "order placement retry trigger," etc. The transaction trigger time can be recorded using a unified clock within the trading server, preferably using a time base consistent with the network card hardware timestamp, the processor's high-precision counter, or the trading program event timestamp, to reduce clock offset in subsequent time window calculations. Trigger conditions can be pre-configured for each branch in the transaction execution tree. These trigger conditions may include one or more of the following: transaction event type matching, order type matching, trading period matching, channel identifier matching, strategy identifier matching, and order placement interface matching. When the current transaction execution information meets the trigger conditions of a certain branch, the processor reads nodes along the corresponding branch starting from the transaction trigger node to form the current target transaction path.

[0120] S1.2: Obtain the node latency characterization information and hardware resource binding relationship of each transaction processing node in the current target transaction path. The node latency characterization information includes at least one of historical execution latency, tail latency ratio and preset latency sensitivity level.

[0121] Specifically, once the target transaction path is determined, the transaction processing nodes along the path do not possess the same low-latency protection value. Some nodes have relatively stable execution latency, and even slight thermal disturbances in their corresponding heat dissipation areas are unlikely to have a perceptible impact on the transaction chain. Other nodes, while having low average execution latency, frequently appear in long-tail latency, easily becoming sources of jitter in opening, call auction, or sparse order placement scenarios. Determining key transaction nodes solely based on node name or the business stage in which the node is located may easily include nodes that "seem important but actually have stable latency" in the protection scope, while potentially overlooking nodes that "have short average latency but are sensitive to tail latency."

[0122] In this embodiment, historical execution latency can be obtained by recording timestamps at the entry and exit points of the transaction processing node. The timestamps can come from transaction program embeddings, gateway call logs, thread execution tracing information, network card hardware timestamps, or processor high-precision timers. For each transaction processing node, a sliding statistics window can be used to store the most recent execution latencies. The sliding statistics window can be set by the number of times or the time length, for example, storing the most recent 10,000 execution records or storing execution records within the most recent 10 minutes. Historical execution latency can include at least one of the following: average execution latency, median execution latency, 90th percentile execution latency, 99th percentile execution latency, and maximum execution latency. The tail latency percentage can be determined based on the proportion of times the execution latency exceeds the node's baseline latency. The node's baseline latency can be the historical median execution latency or the sum of the historical average execution latency and a preset margin. For example, if a node's execution time exceeds 1.5 times its historical median execution time in 200 out of the last 10,000 executions, its tail latency percentage can be recorded as 2%. When the transaction scenario is more sensitive to tail latency, execution records exceeding the 99th percentile or exceeding a preset microsecond threshold can also be included in the tail latency statistics. The preset latency sensitivity level can be configured by the transaction program, for example, setting the order sending node and the rapid order generation node to a high level, the protocol parsing node to a medium level, and the non-critical log processing node to a low level; or the level can be automatically adjusted based on the historical tail latency percentage.

[0123] Furthermore, hardware resource binding relationships are used to locate transaction processing nodes from their software execution location to physical resources within the transaction server. Hardware resource binding relationships can include bindings between thread identifiers and processor cores or core groups, affinity relationships between threads and NUMA nodes, correspondences between threads and memory access regions, interrupt or polling bindings between network interface card (NIC) receive or transmit queues and processor cores, correspondences between order channels and low-latency NIC ports, and connection relationships between PCIe channels and NICs or accelerator cards. These binding relationships can originate from transaction program startup parameters, thread affinity configurations, core binding configurations in the operating environment, NIC queue configurations, IRQ affinity configurations, NUMA policy configurations, or transaction server deployment configurations. For low-latency transaction programs with minimal dynamic thread migration, hardware resource binding relationships can be set to fixed mappings; for programs that may experience thread migration or multi-core collaboration, the main running cores of the node within the statistics window can be recorded, and the core group with the highest runtime percentage can be prioritized for resource binding. For example, if a certain order sending node spends more than 95% of its runtime in the 4th and 5th cores in the last 10,000 executions and consistently uses NIC sending queue 2, then the hardware resource binding relationship of this node can be recorded as "core group 4-5, NIC sending queue 2, corresponding PCIe channel".

[0124] S1.3: The transaction processing node that meets the preset key judgment conditions in the current target transaction path and is mapped to the corresponding liquid cooling heat dissipation node in the heat flow propagation tree through the hardware resource binding relationship is determined as the key transaction node;

[0125] Specifically, determining key transaction nodes involves not simply screening nodes in the target transaction path, but rather a joint assessment of "latency sensitivity" and "heat dissipation controllability." If a transaction processing node has high latency sensitivity but cannot be mapped to a specific liquid-cooled node through hardware resource binding relationships, it is unlikely to be used as a comparison target for subsequent thermal disturbance time windows. Conversely, if a node can be mapped to a liquid-cooled node, but its execution duration is stable and it does not belong to the low-latency protection phase, including it in the thermal flow conflict assessment will broaden the control scope. Therefore, the preset key judgment conditions should not rely on a single indicator, but should be comprehensively judged by combining the transaction stage in which the node is located, the node's latency characterization information, and the liquid-cooling mapping relationship of the node's corresponding hardware resources.

[0126] In one example, the key decision condition includes at least one of the following:

[0127] The transaction processing node belongs to a preset low-latency transaction stage in the current target transaction path, or the transaction processing node is marked as a low-latency protection node under the transaction event type;

[0128] The historical execution latency, tail latency ratio, or preset latency sensitivity level of the transaction processing node meets the corresponding latency sensitivity threshold.

[0129] Specifically, key criteria are used to filter out nodes from multiple transaction processing nodes in the current target transaction path that have protection value for low-latency operation. Since the transaction processing nodes in the transaction server include nodes that directly affect the timing of the transaction chain, such as market data reception, protocol parsing, order generation, and order sending, they may also include nodes that have a weaker impact on critical transaction latency, such as log recording, state synchronization, and non-critical cache updates. Without setting key criteria, subsequent hot flow timing conflict judgments could easily include irrelevant nodes in the analysis scope. Filtering nodes through key criteria allows critical transaction nodes to be concentrated on processing links that are truly latency-sensitive in the current transaction event or are explicitly protected by business configurations, thereby reducing unnecessary calculations in subsequent hot disturbance time window calculations, source tracing, and cooling path adjustments.

[0130] In this embodiment, the transaction processing node belongs to a preset low-latency transaction stage in the current target transaction path. This can be understood as the processing stage corresponding to this node being located in the critical execution section of the low-latency transaction link. The preset low-latency transaction stage can be pre-configured according to the transaction business carried by the transaction server. For example, at least one of the following stages can be configured as a low-latency transaction stage: market data reception stage, protocol parsing stage, strategy response stage, order generation stage, and order placement stage. Different configurations of the preset low-latency transaction stage can be used for different transaction event types. For example, in a sudden market data response event, the market data reception node and protocol parsing node can be configured as low-latency transaction stages; in a call auction or opening order placement event, the order generation node and order placement node can be configured as low-latency transaction stages; in a sparse order placement event, the strategy response node, order generation node, or order placement node that is triggered for the first time after a long period of inactivity can also be configured as a low-latency transaction stage. By configuring low-latency transaction stages according to transaction event types, it is possible to avoid using fixed business nodes as the sole critical nodes in all scenarios, allowing the selection of critical nodes to change according to the actual processing focus of the current transaction event.

[0131] In another example, the historical execution latency of the transaction processing node meets the corresponding latency sensitivity threshold, which can be determined using an absolute threshold, a relative threshold, or a percentile threshold. The absolute threshold can be configured according to the latency requirements of the transaction server for a certain stage. For example, the historical execution latency threshold for the order sending node can be set to 10 microseconds, and the historical execution latency threshold for the protocol parsing node can be set to 5 microseconds. When the node's average execution latency, median execution latency, 90th percentile execution latency, or 99th percentile execution latency within the statistical window exceeds the corresponding threshold, the node is considered to meet the latency sensitivity condition. The relative threshold can be determined based on the node's own historical benchmark. For example, using the median execution latency of the node during a stable operation phase as a benchmark, when the 90th percentile execution latency within the current statistical window exceeds 120% or 150% of this benchmark, the node is considered to meet the latency sensitivity condition. Percentile thresholds are suitable for low-latency scenarios where long-tail jitter is of greater concern. For example, when the 99th percentile execution latency of a node exceeds a preset protection value, or when the difference between the 99th percentile execution latency and the median execution latency exceeds a preset difference threshold, that node is considered a candidate for a key transaction node.

[0132] In this embodiment, the statistical window for historical execution latency can be selected based on the trading scenario. For trading scenarios with concentrated timeframes and drastic load changes, such as opening and call auctions, the statistical window can be set to data within the most recent preset time period, such as the most recent 30 seconds, the most recent 1 minute, or the most recent 5 minutes. For ordinary continuous trading scenarios, the statistical window can be set to the most recent preset number of trading events, such as the most recent 1000, 5000, or 10000 similar trading events. If sufficient historical data is lacking in the initial deployment of the trading server, stress test data, replay data, or preset benchmark data can be used as the initial latency benchmark, and gradually updated with actual trading data during operation. With this setting, the latency sensitivity threshold can meet the availability requirements in the initial deployment phase while also being adjusted according to the actual operating status of the trading server.

[0133] S1.4: Determine the execution time window of the key transaction node based on the call order of the current target transaction path, the transaction trigger time, and the historical execution delay of the transaction processing node located before the key transaction node;

[0134] Specifically, the execution time window of a key transaction node needs to represent the expected execution time range of that node within the current transaction event. This time range is not a fixed configuration value, nor is it a simple copy of the node's own historical execution time. Instead, it is calculated by combining the transaction trigger time, the call order of the current target transaction path, and the execution time of preceding nodes. For thermal flow timing conflict judgment, what the key transaction node really needs to compare is "when the thermal disturbance reaches its corresponding liquid cooling node" and "when the key transaction node is in the execution state." If the execution time window is too narrow, the impact of thermal disturbances during actual execution jitter may be missed; if the execution time window is too wide, irrelevant thermal disturbances may be included in the conflict judgment. Therefore, the determination of the execution time window needs to be processed in conjunction with the node's historical execution data and protection margin, so that it can both cover the actual runtime of the key transaction node and avoid excessively expanding the scope of timing conflict judgment.

[0135] In this embodiment, the processor can read all preceding transaction processing nodes located before the key transaction node from the current target transaction path, and sequentially add the historical execution delays of the preceding nodes according to the calling order to obtain the expected arrival time of the key transaction node relative to the transaction triggering time. The historical execution delay can be selected from one or more of the following: average execution delay, median execution delay, 90th percentile execution delay, or 99th percentile execution delay. For stable preceding nodes, the median execution delay can be used as the estimation basis; for preceding nodes with large tail delay fluctuations, the 90th percentile or 99th percentile execution delay can be used as the estimation basis to avoid premature estimation of the key node execution time due to preceding node jitter. For example, a target transaction path includes a market data receiving node, a protocol parsing node, an order generation node, and an order sending node, with the transaction trigger time being T0. If the 90th percentile execution latency of the market data receiving node is 3 microseconds, the 90th percentile execution latency of the protocol parsing node is 5 microseconds, and the 90th percentile execution latency of the order generation node is 4 microseconds, then the expected execution start time of the order sending node can be determined based on its position approximately 12 microseconds after T0, and combined with the historical execution latency of the order sending node itself, the expected execution end time can be obtained. The above example is only used to illustrate the time window calculation method; the actual execution latency can be determined based on the measurement accuracy of the transaction server and the characteristics of the transaction chain.

[0136] In one example, the determination of thermal disturbance states can be achieved by establishing separate thermal flow benchmarks for each liquid cooling node. Different liquid cooling nodes correspond to different heat dissipation areas, resulting in variations in heat capacity, heat transfer efficiency, and coupling with hardware resources. For instance, the processor cold plate area is more sensitive to power consumption changes, the low-latency network card heat dissipation area is more sensitive to local temperature rises and branch flow changes, and the power supply area is more susceptible to continuous power consumption increases. Therefore, the steady-state thermal flow benchmark should not use a uniform temperature threshold for the entire system, but rather be determined separately based on the node's heat dissipation area, current liquid cooling inlet temperature, branch flow rate, pump speed, coolant temperature difference, and the basic power consumption range of the corresponding hardware resources. In practice, under stable load or off-peak trading conditions, temperature change rate, power consumption change rate, coolant temperature difference, and flow rate data can be continuously collected for several sampling periods for each liquid cooling node. When the node temperature change rate remains within a small fluctuation range within a preset time, for example, the temperature change rate does not exceed ±0.03℃ / s within 60 consecutive seconds, and the corresponding hardware resource power consumption fluctuation does not exceed 5% of the base power consumption, and the branch flow rate fluctuation does not exceed 3% of the rated flow rate, the node thermal state at this time can be recorded as the steady-state thermal flow benchmark under the current liquid cooling condition. For situations where the inlet liquid temperature, pump speed, or branch flow rate changes, the benchmark data under adjacent operating conditions can be re-matched, or the stored benchmark data can be interpolated and corrected to ensure that subsequent thermal disturbance identification is not affected by the switching of liquid cooling operating modes.

[0137] During operation, the processor determines the deviation of each node from the steady-state heat flow benchmark based on the real-time node thermal status data collected from each liquid-cooled heat dissipation node. In the node thermal status data, the node temperature change rate characterizes the heat accumulation trend in the heat dissipation area, the corresponding hardware resource power consumption change rate characterizes the change in heat source input, the coolant temperature difference change reflects the change in heat carried away by the coolant, and the coolant flow rate is used to correct the current heat dissipation capacity. All of the above data can be read within the same sliding sampling window. The length of the sampling window can be set according to the thermal response speed of the trading server and the time scale of trading events; for example, 500 milliseconds to 1 second during normal trading, and 50 milliseconds to 200 milliseconds during opening order placement, call auction, or sudden market events. If the temperature change rate of a processor cold plate node is 0.02℃ / s and the coolant temperature difference is 1.4℃ under steady-state thermal flux reference, and the temperature change rate increases to 0.16℃ / s within the current sampling window, while the corresponding processor core power consumption increases by more than 30W within 200 milliseconds and the coolant temperature difference increases to 2.1℃, then the node can be considered to have deviated from the steady-state thermal flux reference. Here, the focus is on the trend of the node's thermal state relative to the steady-state reference, rather than the absolute magnitude of the node temperature, thus enabling the identification of propagating non-steady-state thermal disturbances before the temperature reaches the traditional alarm threshold.

[0138] Once the thermal flow deviation information meets the disturbance judgment conditions, the corresponding liquid cooling heat dissipation node is recorded as a candidate thermal disturbance node, and a thermal disturbance state is generated for subsequent propagation and calculation. The disturbance judgment conditions can be determined by combining the node's historical steady-state fluctuations to avoid misjudgment caused by sensor noise or short-term power consumption spikes. For example, the node temperature change rate exceeding the mean steady-state temperature change rate plus 3 times the steady-state dispersion, or the corresponding hardware resource power consumption change rate exceeding the mean steady-state power consumption change rate plus 2 times the steady-state dispersion, or the coolant temperature difference change exceeding 20% ​​of the steady-state temperature difference can be used as candidate conditions for entering the disturbance state. To ensure the continuity of the thermal disturbance, the above conditions can be required to be met continuously for no less than two sampling periods. If the mean steady-state temperature change rate of a node is 0.02℃ / s and the steady-state dispersion is 0.03℃ / s, then the temperature change rate entry threshold can be set to 0.11℃ / s; when the current temperature change rate is higher than 0.11℃ / s for two consecutive sampling periods, the sampling time when it first exceeds the entry threshold is taken as the time when the thermal disturbance occurs. The intensity of thermal disturbance can be determined by considering the excess of temperature change rate, excess of power consumption change rate, coolant temperature difference change, and the change in current flow rate relative to the reference flow rate. A lower flow rate than the reference flow rate increases the evaluation of thermal disturbance intensity, while a higher flow rate decreases it. The duration of the thermal disturbance is calculated from the first time the thermal flow deviation information meets the disturbance judgment condition until the thermal flow deviation information falls below the exit condition. The exit condition can be lower than the entry condition; for example, the steady-state mean plus 1.5 times the steady-state dispersion can be used as the exit threshold, requiring the thermal disturbance recording to end after three consecutive sampling periods below the exit threshold. Separating the entry and exit thresholds reduces the repeated generation and disappearance of thermal disturbance states caused by boundary fluctuations, ensuring the continuity of subsequent thermal disturbance time window calculations.

[0139] Next, we will further elaborate on the technical content of the thermal disturbance time window in this application.

[0140] It should be noted that the thermal disturbance time window in this application can be understood as: the expected time range for a thermal disturbance generated at a candidate thermal disturbance node to reach the corresponding liquid cooling node of the key transaction node after propagating along one or more effective propagation paths in the heat flow propagation tree. The thermal disturbance time window is not the original time when the sensor collects the temperature change, nor is it the duration of the local temperature anomaly at the candidate thermal disturbance node. Rather, it is the arrival time range obtained after comprehensively processing the time of thermal disturbance generation, duration, propagation path, propagation delay, and propagation attenuation. Only by converting the thermal disturbance state into the arrival time range at the target liquid cooling node can it be subsequently compared with the execution time window of the key transaction node. For those skilled in the art, the determination of the thermal disturbance time window can be achieved using methods based on flow path calibration, thermal response testing, node thermal capacity estimation, or historical thermal response learning. This application provides the following determination method as an example:

[0141] Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the target liquid cooling heat dissipation node corresponding to the key transaction node is determined;

[0142] In the heat flow propagation tree, the heat flow propagation path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node is determined. For a single-path structure, the parent-child node sequence between the candidate thermal disturbance node and the target liquid cooling heat dissipation node can be directly read along the coolant flow direction. For liquid cooling heat dissipation loops with parallel branches, merging nodes, or bypass paths, effective propagation paths can be selected based on the current branch flow rate, valve status, bypass status, and liquid cooling operating mode. For example, if the current flow rate of a parallel branch is less than 5% of the rated flow rate, or the valve opening is less than a preset effective opening, such as 10%, the branch can be marked as a low-effective path and not used as the main propagation path for the current thermal disturbance time window calculation. If there are multiple effective propagation paths between the candidate thermal disturbance node and the target liquid cooling heat dissipation node, the arrival time window corresponding to each propagation path is calculated separately, and subsequent selection of retention or merging can be based on arrival intensity and time window overlap.

[0143] Based on the flow path length between adjacent liquid cooling nodes in the heat flow propagation path, the current coolant flow rate, and the thermal disturbance attenuation information, the cumulative propagation delay of the thermal disturbance from the candidate thermal disturbance node to the target liquid cooling node is determined.

[0144] Based on the thermal disturbance state and the cumulative propagation delay, the thermal disturbance time window of the target liquid cooling heat dissipation node is determined;

[0145] Specifically, the thermal disturbance state includes the time of thermal disturbance occurrence, the intensity of thermal disturbance, and the duration of thermal disturbance. The processor shifts the time of thermal disturbance occurrence according to the cumulative propagation delay and synchronously shifts the duration of thermal disturbance to the target liquid cooling node, forming the initial arrival time range. Considering flow fluctuations, sensor sampling errors, thermal response calibration errors, and differences in thermal capacity among different nodes, the initial arrival time range can also have an additional time margin. The time margin can be set to a fixed value, such as 2 milliseconds, 5 milliseconds, or 10 milliseconds; or it can be set according to a certain percentage of the cumulative propagation delay, such as 5% to 10% of the cumulative propagation delay. When the arrival intensity of the candidate thermal disturbance node after attenuation along the propagation path is lower than the preset effective disturbance threshold, a corresponding thermal disturbance time window may not be generated; when the arrival intensity meets the effective disturbance threshold, the arrival time range expanded by the time margin is determined as the thermal disturbance time window of the target liquid cooling node. If the arrival time ranges of multiple candidate thermal disturbance nodes overlap at the target liquid cooling heat dissipation node, multiple thermal disturbance time windows can be retained for subsequent conflict contribution judgment, or a superimposed time window can be formed in the overlapping part, and the arrival intensity within the superimposed time window can be merged and evaluated.

[0146] It should be noted that the mapping relationship can be established through hardware resource binding relationships and heat dissipation area coverage relationships. Specifically, the key transaction nodes in the transaction execution tree typically correspond to one or more hardware resources, such as processor core groups, low-latency network card queues, PCIe channels, memory access areas, or order interfaces; the liquid cooling nodes in the heat flow propagation tree correspond to processor cold plate areas, network card heat dissipation areas, memory heat dissipation areas, power supply heat dissipation areas, liquid cooling tributary nodes, or merge nodes. The processor can first determine the hardware resources that the key transaction nodes depend on based on the thread identifier, network card queue identifier, or hardware resource binding relationship of the key transaction nodes, and then determine the target liquid cooling node according to the preset "hardware resource - heat dissipation area - liquid cooling node" correspondence table. For example, when an order sending node is bound to a low-latency network card sending queue, the liquid cooling node corresponding to the network card heat dissipation area can be determined as the target liquid cooling node; when a protocol parsing node is bound to a specific processor core group, the processor cold plate area covering that core group can be determined as the target liquid cooling node. If a key node in a transaction is mapped to multiple liquid cooling nodes, the primary target liquid cooling node can be determined based on the proportion of hardware resources in operation, the node's latency sensitivity level, the thermal coupling strength of the heat dissipation area, or the correlation of historical tail delay. The remaining nodes can be used as auxiliary target liquid cooling nodes to calculate the thermal disturbance time window.

[0147] Additionally, it should be noted that the flow path length, current coolant flow rate, and thermal disturbance attenuation information can be determined jointly using configuration data, sensor data, and calibration data. The flow path length can be obtained from the liquid cooling circuit design drawings, assembly parameters, pipe calibration data, or internal flow channel parameters of the cold plate, or it can be written into the configuration file during server factory testing. The current coolant flow rate can be directly collected from the branch flow meter, or it can be estimated based on pump speed, valve opening, pipe resistance, and total flow rate. For branches without independent flow meters, a "pump speed - valve opening - branch flow rate" correspondence table obtained from factory calibration can be used for lookup. Thermal disturbance attenuation information can be determined through thermal response calibration. For example, during the trading server debugging phase, a step power load is applied to a certain hardware resource, and the temperature changes, coolant temperature difference changes, and response times of the corresponding liquid cooling nodes and downstream nodes are recorded. This yields the disturbance attenuation ratio and response hysteresis between adjacent liquid cooling nodes. For example, after applying a step load to the processor core assembly for 30 seconds, if the processor cold plate node shows a significant heat flow deviation, and the peak temperature change of the downstream network card cooling node is approximately 40% of the peak value of the upstream node, with a time lag of about 0.8 seconds, then the thermal disturbance attenuation information and propagation hysteresis of this adjacent path can be written into the edge attributes of the heat flow propagation tree. Calibration tables can be established for different pump speeds and different inlet liquid temperatures. During runtime, adjacent calibration data can be selected or interpolation corrections can be performed based on the current liquid cooling conditions.

[0148] It should be noted that the cumulative propagation delay can be obtained by summing the edge delays between adjacent liquid cooling nodes in the heat flow propagation path segment by segment. The edge delay between adjacent nodes can be characterized by the residence time of the coolant along the flow path, the thermal inertia hysteresis of the cold plate or heat dissipation area, and the local thermal coupling hysteresis. For the main coolant flow path, the edge delay can be determined based on the effective flow path volume and the current coolant flow rate, and corrected by the response hysteresis caused by the node's thermal capacity. For the internal flow channel of the cold plate or the short-distance thermal coupling area, the hysteresis time obtained from the factory thermal response test can be used as the edge delay. For example, when the equivalent flow path length between adjacent nodes is 0.5 meters, the inner diameter of the flow channel is about 6 millimeters, and the current flow rate is about 1.2 liters / minute, the residence time of the coolant in this flow path is about 0.7 seconds. If the cold plate area is calibrated to have a thermal response hysteresis of about 0.2 seconds, then the edge delay between adjacent nodes can be accumulated at about 0.9 seconds. In the case where there are multiple adjacent nodes in the same path, the processor sequentially accumulates the delay of each edge to obtain the cumulative propagation delay from the candidate thermal disturbance node to the target liquid cooling node; for multiple valid propagation paths, multiple cumulative propagation delays are obtained respectively.

[0149] Next, we will further elaborate on the technical aspects of the method in this application regarding the thermal disturbance source node and the adjustable cooling path.

[0150] Understandably, even after determining that the thermal disturbance time window overlaps with the execution time window of the key transaction node, the liquid cooling node with the highest temperature or the upstream node closest to the target liquid cooling node cannot be directly identified as the source node of the thermal disturbance. Thermal disturbances in low-latency transaction servers exhibit clear temporal and propagation characteristics. A high temperature at a liquid cooling node may originate from continuous heat generation under stable load and may not necessarily participate in the heat flow superposition within the key transaction time window. Conversely, some nodes may not have high absolute temperature values, but if the time of their thermal disturbance generation, duration, and propagation path coincide with the execution time window of the key transaction node, they are more likely to become the actual source of the heat flow timing conflict. Therefore, the source node of the thermal disturbance described in this application is not simply determined by temperature amplitude, but rather refers to a candidate thermal disturbance node that can explain the conflict time window at the target liquid cooling node in terms of time, path, and intensity.

[0151] In one example, the determination of the thermal disturbance source node and the adjustable cooling path includes:

[0152] S3.1: When the thermal disturbance time window and the execution time window meet the preset timing conflict condition, the overlapping part of the thermal disturbance time window and the execution time window is determined as the conflict time window;

[0153] Specifically, the timing conflict condition is used to determine whether the time range of a thermal disturbance arriving at the target liquid cooling node falls within the actual execution range of the transaction critical node. Since the execution time window of the transaction critical node reflects the time period during which the transaction processing node is in a low-latency sensitive execution state, while the thermal disturbance time window reflects the time period during which the candidate thermal disturbance propagates to the target liquid cooling node, when the two overlap in time, it indicates that the thermal disturbance may change the thermal margin of the corresponding heat dissipation area during the execution of the transaction critical node.

[0154] In this embodiment, the timing conflict condition can be determined jointly based on the overlap duration, overlap ratio, and thermal disturbance arrival intensity. The processor can read the execution time window and thermal disturbance time window of the key transaction node. If there is an intersection between the two, a candidate overlap interval is obtained first. Then, it is determined whether the duration of the candidate overlap interval reaches the preset minimum conflict duration, or whether the proportion of the candidate overlap interval to the execution time window of the key transaction node reaches the preset overlap ratio. For example, for ordinary continuous transaction events, the minimum conflict duration can be set to 10% to 20% of the median historical execution duration of the key transaction node, or set to 1 microsecond to 5 microseconds; for opening order placement, call auction, or sparse order placement events, the minimum conflict duration can be reduced to 0.5 microseconds to 2 microseconds. The overlap ratio can be set to 5%, 10%, or 20%, and can be adjusted according to the latency sensitivity level of the key transaction node. The higher the latency sensitivity level, the lower the overlap ratio threshold. If the thermal disturbance arrival intensity in the candidate overlap interval is lower than the effective disturbance threshold, the overlap can be considered not to constitute a valid timing conflict.

[0155] S3.2: Taking the target liquid cooling heat dissipation node as the backtracking starting point, perform conflict time window back projection on each candidate thermal disturbance node along the upstream direction of the heat flow propagation tree to obtain the source matching time window corresponding to each candidate thermal disturbance node;

[0156] Specifically, identifying the source of thermal disturbances cannot stop at the observed conflict at the target liquid cooling node. Instead, it requires searching upstream along the heat flow propagation tree for candidate thermal disturbance nodes that can temporally explain the conflict time window. Since the propagation of thermal disturbances from candidate nodes to the target liquid cooling node involves a path propagation delay, the time of conflict at the target location is not the same as the time of thermal disturbance generation at the source location. The meaning of conflict time window back projection is to push the conflict time range at the target liquid cooling node upstream according to the propagation delay of the candidate path, thereby obtaining the matching time range within which thermal disturbances should occur at the candidate thermal disturbance nodes. Only when a candidate thermal disturbance node actually generates thermal disturbances within the corresponding source matching time window can it possibly explain the temporal conflict at the target location.

[0157] In this embodiment, the processor can start from the target liquid cooling node and backtrack step by step along the heat flow propagation tree towards the coolant inlet or the upstream direction where the thermal disturbance can propagate, reading the heat flow propagation path between each upstream liquid cooling node and the target liquid cooling node. For each candidate thermal disturbance node, the processor accumulates the propagation delays of each edge on the path to obtain the cumulative propagation delay from the candidate node to the target node, and then shifts the entire conflict time window forward by the cumulative propagation delay to obtain the source matching time window. For example, the conflict time window at the target liquid cooling node is from 20 milliseconds to 26 milliseconds after the transaction triggering time. If the cumulative propagation delay from a candidate thermal disturbance node to the target liquid cooling node is 8 milliseconds, then the source matching time window corresponding to the candidate thermal disturbance node can be determined as from 12 milliseconds to 18 milliseconds after the transaction triggering time. If there is an estimation error in the propagation delay, a backtracking margin can be added at both ends of the source matching time window, for example, by adding 0.5 milliseconds to 2 milliseconds, or by adding 5% of the cumulative propagation delay as the error coverage range.

[0158] S3.3: Based on the thermal disturbance generation time, thermal disturbance duration, thermal disturbance intensity, and thermal disturbance attenuation information of the corresponding heat flow propagation path of each candidate thermal disturbance node, determine whether each candidate thermal disturbance node meets the conflict contribution condition, and remove the candidate thermal disturbance nodes that do not meet the conflict contribution condition and their corresponding upstream branches from the scope of this conflict analysis to obtain the conflict subtree;

[0159] Specifically, although candidate thermal disturbance nodes are included in the analysis range after back projection, it is still necessary to further determine whether they contribute sufficiently to the conflict time window. Some candidate nodes may have short-term thermal state changes near the source matching time window, but the duration is too short, and the intensity reaching the target liquid cooling node after propagation path attenuation is insufficient; some candidate nodes, although having high thermal disturbance intensity, have their generation time staggered from the source matching time window, and cannot form the current conflict time window at the target location. The conflict contribution condition is used to simultaneously constrain time matching, intensity matching, and path reachability, so that the branches retained in the conflict subtree can truly participate in the heat flow superposition at the target liquid cooling node.

[0160] In this embodiment, the conflict contribution condition can include three aspects. First, the local disturbance time range formed by the generation time and duration of the thermal disturbance at the candidate thermal disturbance node overlaps with the source matching time window, and the overlap duration reaches a preset source matching threshold. The source matching threshold can be one sampling period or 10% to 20% of the length of the source matching time window. Second, after the thermal disturbance intensity of the candidate thermal disturbance node is corrected by the attenuation information of the corresponding heat flow propagation path, the arrival intensity at the target liquid cooling heat dissipation node reaches a preset conflict contribution intensity threshold. The conflict contribution intensity threshold can be determined based on the steady-state heat flow benchmark of the target liquid cooling heat dissipation node, for example, it can be 1.1 to 1.3 times the upper limit of steady-state heat flow fluctuation; if the key transaction node is in a high latency sensitivity level, the threshold can be reduced to 1.05 to 1.1 times the upper limit of steady-state heat flow fluctuation. Third, the heat flow propagation path is in an effective state under the current liquid cooling conditions, for example, the flow rate of the corresponding branch of the path is not less than 5% to 10% of the rated flow rate, or the valve opening on the path is not less than 10%. Only when all of the above constraints are met can a candidate thermal disturbance node be considered to have contributed to the current conflict.

[0161] Furthermore, the pruning operation can be performed on a temporary analysis copy of the heat flow propagation tree without altering the underlying structure of the preset heat flow propagation tree. The processor can mark candidate thermal disturbance nodes that do not meet the conflict contribution criteria as invalid source nodes and remove branches extending upstream from these invalid source nodes that do not contain other valid candidate thermal disturbance nodes from the scope of this conflict analysis. If an upstream branch has candidate nodes that meet the criteria but the current candidate node does not meet the criteria, only the local analysis mark corresponding to the invalid node is removed, without severing the valid propagation path upstream.

[0162] S3.4: When at least one liquid cooling heat dissipation node in the conflict subtree satisfies a preset short-term thermal coupling condition with the target liquid cooling heat dissipation node, a temporary thermal coupling branch is generated in the conflict subtree between the liquid cooling heat dissipation node that satisfies the preset short-term thermal coupling condition and the target liquid cooling heat dissipation node, and the thermal disturbance source node is determined based on the conflict subtree after the generation of the temporary thermal coupling branch.

[0163] Specifically, the heat flow propagation tree is constructed based on the coolant flow direction, representing the upstream, downstream, and confluence relationships along the main coolant flow path. However, thermal disturbances within the transaction server do not necessarily propagate only along the main coolant flow path. Shared cold plates, metal support components, adjacent PCIe areas, areas near power supply modules, heat recirculation at confluence nodes, or localized airflow recirculation can all cause short-term thermal coupling between two liquid cooling nodes that do not have a direct parent-child relationship in the basic heat flow propagation tree. If the source is attributed solely based on the basic heat flow propagation tree, some laterally transmitted thermal disturbances may not be included in the source range, causing subsequent control paths to only adjust the main flow path while ignoring the coupling sources that actually affect the target liquid cooling node.

[0164] In this embodiment, the preset short-term thermal coupling condition can be determined based on both the physical coupling basis and operational correlation. The physical coupling basis may include two liquid cooling nodes sharing the same cold plate, sharing the same metal heat conductor, being connected to the same bus node, having an installation distance less than a preset spatial distance, being located in the same low-latency network card and PCIe adjacent area, or one node being located within the local heat return influence range of another node. The preset spatial distance can be determined based on the internal structure of the server, for example, 20 mm to 80 mm. The operational correlation can be determined by the change relationship of the heat flow deviation information of the two nodes within a preset time difference. For example, if a liquid cooling node experiences heat flow deviation, and the target liquid cooling node shows a local thermal response within 100 milliseconds to 2 seconds with a temperature change rate increase in the same direction, an increase in coolant temperature difference, or no change in power consumption, and the correlation coefficient between the two reaches 0.6 or higher, then the short-term thermal coupling condition can be considered met. For scenarios with short sampling periods, the method of consistent heat flow deviation direction within multiple consecutive sampling windows can also be used instead of correlation coefficient calculation.

[0165] Furthermore, temporary thermal coupling branches are generated only within the conflict subtree of this conflict analysis. Essentially, they are auxiliary propagation edges and do not indicate a change in the actual piping of the liquid cooling circuit. Temporary thermal coupling branches can be associated with temporary propagation delay, temporary attenuation information, and an effective time period. The temporary propagation delay can be determined based on the time lag of the correlation between the thermal responses of two nodes, for example, by taking the time difference between the peak deviation of heat flow at one node and the peak response of the target node. The temporary attenuation information can be determined based on the ratio between the target node's response amplitude and the source node's deviation amplitude. The effective time period can be limited to the conflict analysis period under the current transaction event or the current liquid cooling condition.

[0166] In yet another example, the method for determining the source node of the thermal disturbance based on the conflict subtree after generating the temporary thermally coupled branch includes:

[0167] For each candidate thermal disturbance node in the conflict subtree, an effective arrival path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node is determined. The effective arrival path includes an intra-tree propagation path formed along the heat flow propagation tree, or a temporary coupling path formed via the temporary thermal coupling branch.

[0168] Based on the thermal disturbance state of the candidate thermal disturbance node and the propagation delay and disturbance attenuation information corresponding to the effective arrival path, the path arrival time window and path arrival intensity of the candidate thermal disturbance node reaching the target liquid cooling heat dissipation node via each effective arrival path are determined.

[0169] Based on the degree of overlap between the path arrival time window and the conflict time window, and the path arrival intensity, the path conflict contribution value of the candidate thermal disturbance node via each effective arrival path is determined.

[0170] Multiple path conflict contribution values ​​belonging to the same candidate hot disturbance node are fused to obtain the node conflict contribution value of the candidate hot disturbance node relative to the conflict time window;

[0171] Based on the node conflict contribution value of each candidate thermal disturbance node, a set of source nodes for interpreting the conflict time window is determined from the conflict subtree, and the candidate thermal disturbance nodes in the source node set are determined as the thermal disturbance source nodes.

[0172] Understandably, generating a conflict subtree after creating a temporary thermal coupling branch is equivalent to a local reconstruction of the propagation relationships involved in this heat flow timing conflict. The reconstructed conflict subtree retains the intra-tree propagation path on the main coolant flow path and also allows the introduction of temporary coupling paths formed by shared cold plates, merging nodes, adjacent heat-conducting structures, or short-term thermal correlations. When determining the effective arrival path from a candidate thermal disturbance node to the target liquid-cooled heat dissipation node, the processor can start from the candidate thermal disturbance node, search along the conflict subtree for paths that can reach the target liquid-cooled heat dissipation node, and restrict the effectiveness of the paths. Specifically, the basic condition can be that each edge in the path is in an effective state under the current liquid cooling conditions. For example, the flow rate of the branch corresponding to the intra-tree propagation path is not less than 5% to 10% of the rated flow rate, the short-term thermal coupling branch corresponding to the temporary coupling path is still within the effective time period, and the correlation or response strength of the temporary thermal coupling branch is not lower than the exit threshold at the time of generation. For temporary thermal coupling branches that may form loops, it can be restricted that liquid cooling nodes in each effective arrival path cannot appear repeatedly, or that temporary thermal coupling branches can be used at most once in the same path. This ensures that the temporary coupling relationship serves only as an auxiliary propagation edge for this conflict attribution, without changing the physical meaning of the basic heat flow propagation tree. The resulting effective arrival paths can cover the conventional propagation in the direction of coolant flow, as well as short-term lateral heat transfer outside the main liquid cooling path, so that the source attribution is not limited to a single flow path.

[0173] For each valid arrival path, the path arrival time window can be determined based on the thermal disturbance generation time of the candidate thermal disturbance node, the duration of the thermal disturbance, and the propagation delay of the valid arrival path. The propagation delay of the intra-tree propagation path can be obtained by summing the flow path residence time between adjacent liquid cooling heat dissipation nodes in the path, the thermal inertia hysteresis of the cold plate, and the node response hysteresis; the propagation delay of the temporary coupling path can be given by the thermal response lag time determined when generating the temporary thermal coupling branch, such as the time difference between the peak deviation of the source node's heat flux and the peak response of the target node. The path arrival intensity can be corrected segment by segment based on the thermal disturbance intensity of the candidate thermal disturbance node according to the disturbance attenuation information in the valid arrival path; for the intra-tree propagation path, the attenuation information can come from the calibration data of branch heat transfer capacity, node heat capacity, and coolant temperature difference change; for the temporary coupling path, the attenuation information can be determined by the ratio between the response amplitude of the target liquid cooling heat dissipation node and the deviation amplitude of the source node. If the time windows of the same candidate thermal disturbance node reaching the target liquid cooling heat dissipation node via different effective arrival paths are close to each other, they can be retained separately for contribution evaluation. If the center interval of the arrival time windows of two paths is less than one sampling period and the paths have a common upstream segment, the path with lower arrival intensity or lower path confidence can be downweighted to avoid the same thermal disturbance being repeatedly amplified due to multiple path expressions.

[0174] The path conflict contribution value can be determined comprehensively based on the overlap between the path arrival time window and the conflict time window, the path arrival intensity, and the path reliability. The overlap can be characterized by the duration of the path arrival time window falling into the conflict time window, the overlap ratio, or the offset of the overlap center; the path arrival intensity can be characterized by the intensity of heat flow deviation after reaching the target liquid cooling heat dissipation node; the path reliability can be determined according to the path type and calibration quality. For example, propagation paths within the basic tree can adopt a higher reliability, while the reliability of temporary coupling paths can be determined based on the correlation of short-term thermal coupling branches, the effective time period, and the number of historical recurrences. For example, if the overlap ratio between the arrival time window and the conflict time window of a certain path reaches 60%, and the arrival intensity exceeds 1.2 times the upper limit of the steady-state fluctuation of the target liquid cooling heat dissipation node, then the path can be assigned a higher path conflict contribution value; if another path has a high arrival intensity, but its arrival time window only overlaps with the edge of the conflict time window for a very short time, then its path conflict contribution value can be reduced accordingly. Multiple path conflict contribution values ​​belonging to the same candidate thermal disturbance node can be fused using methods such as retaining the maximum contribution value, weighted summation, or main path plus auxiliary path compensation. When multiple paths are independent and have different arrival time windows, weighted summation can be used. When multiple paths share the main propagation segment and their arrival time windows are close, it is advisable to use the maximum contribution value and add a smaller compensation to reduce duplicate inclusion. After the node conflict contribution value is determined, candidate thermal disturbance nodes with contribution values ​​higher than a preset source threshold can be included in the source node set. The source threshold can be 30% to 50% of the maximum contribution value of all candidate nodes, or it can be set to ensure that the cumulative contribution of the source node set reaches more than 80% of the total contribution. The thermal disturbance source nodes determined in this way can form a correspondence with the conflict time window in terms of time, intensity, and propagation path, and subsequent liquid cooling adjustment can be carried out around the nodes that actually participate in this heat flow superposition.

[0175] S3.5: Among the paths from the thermal disturbance source node to the target liquid cooling heat dissipation node and the path corresponding to the temporary thermal coupling branch, determine the candidate cooling path containing the adjustable component;

[0176] Specifically, after identifying the source node of the thermal disturbance, it is also necessary to identify which cooling paths can effectively regulate the propagation state of the thermal disturbance. Adjustable components in the liquid cooling circuit may include pumps, branch valves, bypass valves, radiator fans, cooling capacity grading units, electronic expansion valves, or actuators that can change local flow rates and heat exchange capacity. However, not all paths containing adjustable components are suitable as candidate cooling paths. Candidate cooling paths should at least simultaneously meet the following conditions: connected to the source node of the thermal disturbance, able to affect the propagation delay or arrival intensity of the thermal disturbance to the target liquid cooling node, and the adjustment action being effective within the critical timescale of the transaction.

[0177] In this embodiment, the processor can read the adjustable identifiers of each edge and node along the tree-based propagation path from the source node of the thermal disturbance to the target liquid-cooled heat dissipation node. If the opening of the branch valve corresponding to a certain edge is adjustable, or the flow rate of the cold plate branch corresponding to a certain node is adjustable, the propagation path segment containing that edge or node is marked as a candidate cooling path. For paths corresponding to temporary thermal coupling branches, relevant adjustment components that can weaken the coupling transmission can be further searched, such as branch flow rate adjustment components that reduce the intensity of thermal disturbance upstream of the shared cold plate, bypass valves that change the temperature rise of the confluence node, branch valves that improve the local cooling capacity of the target node, or cooling capacity grading units. A candidate cooling path can be an entire path from the source node to the target node, or a local path segment containing key adjustable components. If the adjustment response time of a certain adjustable component is greater than the remaining time between the execution time window of the key transaction node and the current time, then even if it is physically adjustable, it may not be used as a candidate cooling path for this time, or it may only be used as a pre-cooling control path for subsequent transaction events.

[0178] S3.6: Based on the impact of adjusting the candidate cooling path on the thermal disturbance propagation delay, thermal disturbance arrival intensity, and conflict time window, evaluate the adjustment effect of the candidate cooling path, and remove the candidate cooling path that expands the conflict time window or increases the thermal disturbance arrival intensity.

[0179] Specifically, the adjustability of a candidate cooling path does not necessarily mean that its adjustment direction is beneficial. Increasing the flow rate of a certain branch may reduce the intensity of thermal disturbance, but it may also shorten the propagation delay, allowing the thermal disturbance to reach the target liquid-cooled heat dissipation node earlier and enter the execution time window of the critical transaction node. Decreasing the flow rate of a certain path may delay the arrival time of the thermal disturbance, but it may aggravate the local heat accumulation at the source node. Low-latency transaction scenarios are more concerned with the superposition intensity of thermal disturbances during the critical execution time of the transaction. Therefore, the impact assessment of adjustment needs to consider the changes in propagation delay, arrival intensity, and conflict time window simultaneously, and cannot select the cooling path solely based on the cooling rate.

[0180] In this embodiment, the processor can establish a comparison before and after adjustment for each candidate cooling path. Before adjustment, the baseline arrival time window and baseline arrival intensity of the thermal disturbance are obtained based on the current flow rate, current valve opening, current pump speed, and current attenuation information. After adjustment, the predicted arrival time window and predicted arrival intensity are obtained by estimating the new flow rate, propagation delay, and attenuation information based on the candidate adjustment actions. Candidate adjustment actions may include increasing branch flow rate, decreasing branch flow rate, opening bypass, closing bypass, increasing radiator fan speed, switching cooling capacity level, or pre-cooling. The estimation process can be completed by looking up tables, for example, determining the parameters after adjustment based on the calibration table of "valve opening - branch flow rate - propagation delay - attenuation ratio"; or it can be obtained by interpolation based on adjacent operating condition data recorded during operation. If the overlap time between the adjusted thermal disturbance time window and the execution time window of the key transaction node increases, or the overlap ratio increases, or the predicted arrival intensity within the conflict time window is higher than the arrival intensity before adjustment, then the candidate cooling path is removed. The expansion of the conflict time window in this application can be understood as an increase in the overlap duration, an increase in the overlap ratio, or the emergence of new overlaps in time periods that were not originally overlapping due to the adjustment.

[0181] S3.7: Determine the adjustable cooling path as the candidate cooling path that meets the preset conflict reduction condition among the remaining candidate cooling paths;

[0182] Specifically, after adjustment impact assessment and elimination processing, the remaining candidate cooling paths have excluded those that would expand the conflict time window or increase the conflict intensity. However, it is still necessary to select a path that can meet the conflict reduction requirements as an adjustable cooling path. Preset conflict reduction conditions are used to evaluate the degree to which candidate cooling paths improve the timing conflict. The focus is not on a single temperature decrease, but on the reduction of the cumulative impact of thermal disturbances within the execution time window of the key transaction nodes. Therefore, the determination of the adjustable cooling path can be consistent with the heat flow timing control logic of this application, meaning that the control actions revolve around conflict reduction within the key time window.

[0183] In this embodiment, the conflict reduction condition may include at least one of the following: overlap time reduction condition, arrival intensity reduction condition, and adjustment executable condition. The overlap time reduction condition may be set such that the overlap duration between the adjusted thermal disturbance time window and the execution time window decreases by a preset proportion, for example, by more than 20%. The arrival intensity reduction condition may be set such that the arrival intensity within the adjusted conflict time window decreases by a preset proportion, for example, by 15% to 30%. The adjustment executable condition may be set such that the adjustable components on the candidate cooling path can complete the adjustment within a preset response time, for example, before the start of the execution time window of the transaction critical node, or before the arrival of the next similar transaction event. For transaction critical nodes with high latency sensitivity, the overlap time reduction ratio requirement can be reduced while the execution time requirement can be increased; for ordinary transaction nodes, the reduction ratio requirement can be increased to reduce frequent adjustments. If multiple candidate cooling paths meet the conflict reduction condition, one or more paths can be selected based on the degree of conflict reduction, adjustment cost, impact on other nodes, and the operational stability of the liquid cooling loop.

[0184] In another example, when multiple key transaction nodes exist in the current target transaction path, the execution time window, target liquid cooling node, heat disturbance source node, and candidate cooling path corresponding to each key transaction node can be determined separately. Based on this, thermal flow coupling decoupling processing for multiple key transaction nodes is performed. Multiple key transaction nodes can correspond to different transaction processing stages, such as market data receiving nodes, protocol parsing nodes, order generation nodes, and order sending nodes. They can also correspond to different order channels, different network card queues, or different processor core groups within the same transaction stage. Since multiple key transaction nodes may share the same cold plate, the same liquid cooling branch, the same bus node, or adjacent heat conduction areas in the liquid cooling circuit, a liquid cooling control strategy generated specifically for one key transaction node may change the heat disturbance propagation delay, arrival intensity, or duration at the corresponding liquid cooling nodes of other key transaction nodes. Therefore, this example does not simply superimpose the liquid cooling control strategies corresponding to each key transaction node. Instead, it establishes control influence constraints between key transaction nodes to prevent conflict reduction actions at one key transaction node from creating new heat flow timing conflicts at another key transaction node.

[0185] Specifically, the processor can treat each key transaction node as a node to be protected and establish corresponding node conflict description information for each node. This node conflict description information may include the key transaction node identifier, execution time window, target liquid cooling node, thermal disturbance source node, conflict time window, candidate cooling path, and the adjustment impact evaluation result corresponding to the candidate cooling path. For any candidate cooling path, not only is its conflict reduction effect on the current key transaction node calculated, but it is also determined, based on the heat flow propagation tree, whether the candidate cooling path shares branches, cold plates, common confluence nodes, or temporary thermal coupling branches with the target liquid cooling nodes corresponding to other key transaction nodes. If such sharing relationships exist, the adjustment action of the candidate cooling path is projected onto other key transaction nodes, predicting the changes in the thermal disturbance time window, arrival intensity, and conflict time window of the other key transaction nodes after adjustment. Thus, the control impact relationship of the candidate cooling path relative to multiple key transaction nodes can be obtained, rather than only its local control effect on a single key transaction node.

[0186] In this embodiment, a key node coupling table can be constructed based on the control influence relationship. The key node coupling table can have transaction key nodes as rows and candidate cooling paths as columns, recording the type and magnitude of each candidate cooling path's impact on each transaction key node. The impact type can include conflict reduction, conflict enhancement, time window shift, intensity reduction, intensity increase, or no substantial impact; the impact magnitude can be determined by the change in overlap time before and after adjustment, the change in intensity within the conflict time window, and the center shift of the thermal disturbance time window. For example, if increasing the valve opening of a branch reduces the overlap ratio between the thermal disturbance time window and the execution time window corresponding to the order sending node from 60% to 20%, but simultaneously advances the thermal disturbance time window corresponding to the protocol parsing node and causes a 15% overlap with its execution time window, then this branch adjustment is recorded as conflict reduction for the order sending node and conflict enhancement for the protocol parsing node. Through this key node coupling table, it is possible to identify whether candidate cooling paths have adverse cross-node effects before the control strategy is generated.

[0187] Furthermore, the processor can divide multiple key trading nodes into several thermal-fluid coupling groups based on the key node coupling table. If the target liquid cooling nodes corresponding to two key trading nodes are located downstream of the same liquid cooling branch, or if they share adjustable components in their candidate cooling paths, or if adjusting the candidate cooling path of one key trading node would cause the conflict evaluation value of the other key trading node to increase beyond a preset coupling threshold, then the two are classified into the same thermal-fluid coupling group. The preset coupling threshold can be determined based on the latency sensitivity level of the key trading nodes. For example, if the overlap ratio of the other key trading node increases by more than 10% after adjustment, or if the arrival intensity within the conflict time window increases by more than 10% of the upper limit of steady-state thermal flux fluctuation, then the two can be considered to have a thermal-fluid coupling relationship that needs to be coordinated. For high latency sensitive events such as opening order placement and call auction, the overlap ratio increase threshold can be reduced to 5%. Key trading nodes that do not belong to the same thermal-fluid coupling group can continue to generate liquid cooling control strategies according to the aforementioned single-node processing method; key trading nodes that belong to the same thermal-fluid coupling group enter the decoupling process.

[0188] For the same thermal flux coupling group, the common upstream node, common liquid cooling branch, or common merging node of each transaction key node corresponding to the target liquid cooling heat dissipation node can be determined in the thermal flux propagation tree, thereby determining the common cooling segment and the private cooling segment. The common cooling segment refers to the path segment that, after adjustment, will simultaneously affect two or more transaction key nodes, while the private cooling segment refers to the path segment that mainly affects a single transaction key node. For example, if a branch from the processor cold plate inlet to a certain merging node simultaneously serves both the protocol parsing node and the order generation node, then this branch can be marked as a common cooling segment; a branch after the merging node that only affects the network card heat dissipation area can be marked as the private cooling segment corresponding to the order sending node. By distinguishing between common cooling segments and private cooling segments, the candidate cooling path can be divided into a common impact part and a node-specific impact part, avoiding the uncontrollable shift in the thermal disturbance time window of multiple transaction key nodes caused by prioritizing the adjustment of the common path.

[0189] In this embodiment, virtual decoupling nodes can be generated at the connection points between the common cooling section and each private cooling section. These virtual decoupling nodes do not represent new physical components in the liquid cooling circuit, but rather serve to distinguish between common and private adjustment effects in control calculations. Using these virtual decoupling nodes as boundaries, the processor splits candidate cooling paths spanning multiple key transaction nodes into common candidate paths and private candidate paths. For private candidate paths, if adjustment reduces the conflict evaluation value of the corresponding key transaction node and its impact on the conflict evaluation values ​​of other key transaction nodes is below a preset crosstalk threshold, it is preferentially retained. For common candidate paths, they are only allowed to enter the candidate control set if they do not increase conflict among any key transaction node within the thermal-fluid coupling group, or if the conflict reduction for high-priority key transaction nodes is greater than the adverse impact on low-priority key transaction nodes and can be compensated for by subsequent private paths. Thus, liquid cooling adjustment preferentially applies to node-specific paths, and only when node-specific paths cannot meet the conflict reduction conditions is the common cooling section activated to a limited extent.

[0190] Furthermore, node protection priorities can be set for each key transaction node within the thermal-fluid coupling group. Node protection priorities can be determined based on the transaction event type, node latency sensitivity level, the distance of the execution time window from the current moment, the historical tail latency percentage, and the transaction processing stage. For example, order sending nodes may have higher priority during call auction or opening order events; key transaction nodes with upcoming execution time windows may have higher priority than nodes with later execution time windows; nodes with a high tail latency percentage may have higher priority than nodes with stable historical execution latency. Node protection priorities are not used to directly discard low-priority nodes, but rather to determine the adjustment direction and compensation order when the common cooling segment inevitably has an impact. For high-priority nodes, the control strategy can prioritize shortening their conflict time windows or reducing their arrival intensity; for low-priority nodes, if adversely affected by common path adjustments, compensation is prioritized through their private cooling segments, or postponed to the next control cycle for mitigation.

[0191] In one specific implementation, the processor can first generate single-node candidate control actions for each transaction critical node within the thermal-fluid coupling group, and map each candidate control action to a critical node coupling table. Then, candidate control actions are selected from high to low node protection priority. Each time a candidate control action is selected, the predicted thermal disturbance time window and predicted arrival intensity of all transaction critical nodes within the thermal-fluid coupling group are updated, and the conflict evaluation value of each node is recalculated. If the current candidate control action causes the conflict evaluation value of any unprocessed transaction critical node to rise above the crosstalk threshold, the candidate control action is not executed directly. Instead, a compensation path is searched in the private cooling segment of the corresponding affected transaction critical node. The compensation path can be a candidate cooling path that can delay the arrival time of the thermal disturbance of the affected node, reduce its arrival intensity, or shorten its overlap ratio. If a compensation path exists, the current candidate control action is combined with the compensation path into a cooperative control action; if no compensation path exists, the adjustment magnitude of the current candidate control action is reduced, or an alternative path with a slightly lower reduction effect on high-priority nodes but a smaller impact on other nodes is selected. This gradual update and compensation process can reduce the mutual offsetting and conflict migration between control actions of key transaction nodes.

[0192] For conflict evaluation values, a comprehensive quantity reflecting the degree of temporal conflict can be used. The conflict evaluation value can be determined based on the overlap ratio between the thermal disturbance time window and the execution time window, the arrival intensity within the conflict time window, the latency sensitivity level of key transaction nodes, and the urgency of the execution time window relative to the current moment. For example, when the overlap ratio of a key transaction node decreases by 30%, the arrival intensity decreases by 20%, and it does not cause an increase in the overlap ratio of other nodes exceeding 5%, the corresponding adjustment action can be considered to meet the decoupling control conditions. When the adjustment action reduces the conflict evaluation value of the target node but increases the conflict evaluation value of another high-priority node by more than 10%, the adjustment action is considered to have a conflict migration risk and requires compensation or replacement. For ordinary trading periods, the crosstalk threshold can be taken as 10% to 15%; for call auction, opening order placement, or sparse order placement scenarios, the crosstalk threshold can be taken as 5% to 10%. These thresholds can be jointly set by the historical tail latency tolerance of the trading server, the liquid cooling loop adjustment response time, and the importance of the trading event.

[0193] After selecting and compensating for candidate control actions, the processor can generate a joint liquid cooling control strategy for the thermal-fluid coupling group. This joint liquid cooling control strategy includes the adjustment target, direction, magnitude, and execution time for each adjustable cooling path. For the common cooling section, the adjustment magnitude can be gradually varied, for example, limiting the change in pump speed or valve opening within a control cycle to no more than 5% to 10% of the rated range, to reduce the simultaneous disturbance of multiple key trading nodes by the common adjustment. For the private cooling section, if it primarily affects a single key trading node, faster local adjustment can be allowed to offset or reduce the intensity of thermal disturbances before the execution window of that key trading node arrives. For trading events involving pre-cooling requirements, such as opening order placement or call auction, the joint liquid cooling control strategy can also pre-adjust the corresponding private cooling section before the execution window of the key trading node, ensuring that thermal disturbances are peak-shaving or staggered before the critical execution time.

[0194] Through the decoupling process described above, while multiple key transaction nodes can still independently perform thermal disturbance time window judgment and source tracing, the output of the liquid cooling control strategy is no longer a simple sum of multiple single-node strategies. The common cooling segment, private cooling segment, virtual decoupled nodes, and key node coupling table collectively enable the processor to identify the direction of an adjustment action's impact on multiple key transaction nodes and constrain conflict migration before adjustment. Therefore, even if multiple key transaction nodes share the same liquid cooling branch or adjacent heat dissipation areas, they can coordinate the adjustment actions of different cooling paths while maintaining the independence of their respective timing judgments, making thermal flow balance control more suitable for the multi-node concurrent execution characteristics of low-latency transaction processes.

[0195] In yet another example, embodiments of this application provide a liquid cooling heat dissipation control system based on thermal flux balance, the system comprising:

[0196] The information acquisition module is used to acquire the current transaction execution information and current thermal status information of the transaction server, and to determine the execution time window of the key nodes of the transaction based on the current transaction execution information. The current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node.

[0197] The timing conflict analysis module is used to determine the thermal disturbance time window for thermal disturbance propagation to the liquid cooling node corresponding to the key transaction node based on the preset transaction execution tree, heat flow propagation tree, and mapping relationship between transaction processing nodes and liquid cooling nodes in the transaction server. When the thermal disturbance time window and the execution time window meet the preset timing conflict conditions, the module reversely determines the source node of the thermal disturbance and the adjustable cooling path based on the heat flow propagation tree. The nodes of the transaction execution tree are used to represent transaction processing nodes, and the nodes of the heat flow propagation tree are used to represent liquid cooling nodes.

[0198] The liquid cooling control module is used to generate a liquid cooling heat dissipation control strategy based on the heat disturbance source node and the adjustable cooling path, and to adjust the working state of the liquid cooling heat dissipation circuit of the transaction server according to the liquid cooling heat dissipation control strategy.

[0199] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A liquid cooling heat dissipation control method based on heat flux balance, applied to the processor chip of a trading server, wherein the trading server is configured with a processor and a liquid cooling heat dissipation circuit, characterized in that, The transaction server is pre-configured with a transaction execution tree and a heat flow propagation tree for the processor to call. Nodes in the transaction execution tree represent transaction processing nodes, and nodes in the heat flow propagation tree represent liquid cooling nodes. The method includes: Obtain the current transaction execution information and current thermal status information of the transaction server, and determine the execution time window of the key transaction nodes based on the current transaction execution information, wherein the current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node; Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the heat disturbance time window for the propagation of heat disturbance to the liquid cooling heat dissipation node corresponding to the transaction key node is determined. When the heat disturbance time window and the execution time window meet the preset timing conflict condition, the heat disturbance source node and the adjustable cooling path are determined in reverse based on the heat flow propagation tree. A liquid cooling heat dissipation control strategy is generated based on the thermal disturbance source node and the adjustable cooling path, and the working state of the liquid cooling heat dissipation circuit is adjusted according to the liquid cooling heat dissipation control strategy.

2. The liquid cooling heat dissipation control method based on heat flow balance according to claim 1, characterized in that, The methods for constructing the transaction execution tree include: Obtain the transaction link configuration data of the transaction server. The transaction link configuration data includes the calling order between transaction processing stages, the thread identifier for executing each transaction processing stage, and the hardware resource binding relationship corresponding to the thread identifier. The transaction trigger node is used as the root node of the transaction execution tree, and at least some of the nodes among the market data receiving node, protocol parsing node, strategy response node, order generation node and order sending node are connected as child nodes of the transaction execution tree according to the calling order between the transaction processing stages.

3. The liquid cooling heat dissipation control method based on heat flow balance according to claim 1, characterized in that, The heat flow propagation tree is constructed in the following ways: Obtain the flow path topology data of the liquid cooling heat dissipation circuit, which includes the connection relationship between the coolant inlet, cold plate, liquid cooling branch, junction node, radiator and coolant outlet; The coolant inlet is used as the root node of the heat flow propagation tree, and the liquid cooling nodes corresponding to different heat dissipation areas are connected to form the heat flow propagation tree according to the flow direction of the coolant in the liquid cooling heat dissipation circuit.

4. The liquid cooling heat dissipation control method based on heat flow balance according to claim 2, characterized in that, The methods for determining the key nodes of the transaction include: Based on the transaction event type and transaction trigger time in the current transaction execution information, determine the current target transaction path in the transaction execution tree; Obtain the node latency characterization information and hardware resource binding relationship of each transaction processing node in the current target transaction path. The node latency characterization information includes at least one of historical execution latency, tail latency ratio, and preset latency sensitivity level. The transaction processing node that meets the preset key judgment conditions in the current target transaction path and is mapped to the corresponding liquid cooling heat dissipation node in the heat flow propagation tree through the hardware resource binding relationship is identified as the key transaction node. The execution time window of the key transaction node is determined based on the call order of the current target transaction path, the transaction trigger time, and the historical execution delay of the transaction processing node located before the key transaction node.

5. The liquid cooling heat dissipation control method based on heat flow balance according to claim 4, characterized in that, The key determination criteria include at least one of the following: The transaction processing node belongs to a preset low-latency transaction stage in the current target transaction path, or the transaction processing node is marked as a low-latency protection node under the transaction event type; The historical execution latency, tail latency ratio, or preset latency sensitivity level of the transaction processing node meets the corresponding latency sensitivity threshold.

6. The liquid cooling heat dissipation control method based on heat flow balance according to claim 3, characterized in that, The methods for determining the thermal disturbance state include: For each liquid cooling heat dissipation node in the heat flow propagation tree, the steady-state heat flow reference of the liquid cooling heat dissipation node is determined according to the steady-state thermal state of the heat dissipation area corresponding to the liquid cooling heat dissipation node under the current liquid cooling conditions. Acquire node thermal state data of the heat dissipation area corresponding to the liquid cooling heat dissipation node. The node thermal state data includes at least one of node temperature change rate, corresponding hardware resource power consumption change rate, coolant temperature difference change, and coolant flow rate. Based on the degree of deviation of the node thermal state data from the steady-state heat flow reference, the heat flow deviation information of the liquid cooling heat dissipation node is determined; When the heat flow deviation information meets the preset disturbance judgment conditions, the corresponding liquid cooling heat dissipation node is determined as a candidate thermal disturbance node, and the thermal disturbance state is generated according to the heat flow deviation information of the candidate thermal disturbance node. The thermal disturbance state includes the thermal disturbance generation time, thermal disturbance intensity, and thermal disturbance duration.

7. The liquid cooling heat dissipation control method based on heat flow balance according to claim 6, characterized in that, The method for determining the thermal disturbance time window includes: Based on the mapping relationship between the transaction execution tree and the heat flow propagation tree, the target liquid cooling heat dissipation node corresponding to the key transaction node is determined; In the heat flow propagation tree, determine the heat flow propagation path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node; Based on the flow path length between adjacent liquid cooling nodes in the heat flow propagation path, the current coolant flow rate, and the thermal disturbance attenuation information, the cumulative propagation delay of the thermal disturbance from the candidate thermal disturbance node to the target liquid cooling node is determined. The thermal disturbance time window of the target liquid cooling heat dissipation node is determined based on the thermal disturbance state and the cumulative propagation delay.

8. The liquid cooling heat dissipation control method based on heat flow balance according to claim 7, characterized in that, The methods for determining the thermal disturbance source node and the adjustable cooling path include: When the thermal disturbance time window and the execution time window meet the preset timing conflict conditions, the overlapping part of the thermal disturbance time window and the execution time window is determined as the conflict time window; Taking the target liquid cooling heat dissipation node as the backtracking starting point, the conflict time window of each candidate thermal disturbance node is back-projected along the upstream direction of the heat flow propagation tree to obtain the source matching time window corresponding to each candidate thermal disturbance node. Based on the thermal disturbance generation time, duration, intensity, and thermal disturbance attenuation information of the corresponding heat flow propagation path of each candidate thermal disturbance node, it is determined whether each candidate thermal disturbance node meets the conflict contribution condition. Candidate thermal disturbance nodes that do not meet the conflict contribution condition and their corresponding upstream branches are removed from the scope of this conflict analysis to obtain the conflict subtree. When at least one liquid cooling heat dissipation node in the conflict subtree satisfies a preset short-term thermal coupling condition with the target liquid cooling heat dissipation node, a temporary thermal coupling branch is generated in the conflict subtree between the liquid cooling heat dissipation node that satisfies the preset short-term thermal coupling condition and the target liquid cooling heat dissipation node, and the thermal disturbance source node is determined based on the conflict subtree after the generation of the temporary thermal coupling branch. Among the paths from the thermal disturbance source node to the target liquid cooling heat dissipation node and the paths corresponding to the temporary thermal coupling branches, candidate cooling paths containing adjustable components are determined. Based on the impact of adjusting the candidate cooling path on the thermal disturbance propagation delay, thermal disturbance arrival intensity, and conflict time window, the adjustment impact of the candidate cooling path is evaluated, and candidate cooling paths that increase the conflict time window or increase the thermal disturbance arrival intensity are removed. The candidate cooling path that meets the preset conflict reduction condition among the remaining candidate cooling paths is determined as the adjustable cooling path.

9. The liquid cooling heat dissipation control method based on heat flow balance according to claim 8, characterized in that, The methods for determining the source node of the thermal disturbance based on the conflict subtree after generating the temporary thermally coupled branch include: For each candidate thermal disturbance node in the conflict subtree, an effective arrival path from the candidate thermal disturbance node to the target liquid cooling heat dissipation node is determined. The effective arrival path includes an intra-tree propagation path formed along the heat flow propagation tree, or a temporary coupling path formed via the temporary thermal coupling branch. Based on the thermal disturbance state of the candidate thermal disturbance node and the propagation delay and disturbance attenuation information corresponding to the effective arrival path, the path arrival time window and path arrival intensity of the candidate thermal disturbance node reaching the target liquid cooling heat dissipation node via each effective arrival path are determined. Based on the degree of overlap between the path arrival time window and the conflict time window, and the path arrival intensity, the path conflict contribution value of the candidate thermal disturbance node via each effective arrival path is determined. Multiple path conflict contribution values ​​belonging to the same candidate hot disturbance node are fused to obtain the node conflict contribution value of the candidate hot disturbance node relative to the conflict time window; Based on the node conflict contribution value of each candidate thermal disturbance node, a set of source nodes for interpreting the conflict time window is determined from the conflict subtree, and the candidate thermal disturbance nodes in the source node set are determined as the thermal disturbance source nodes.

10. A liquid cooling heat dissipation control system based on heat flux balance, used to implement the liquid cooling heat dissipation control method based on heat flux balance as described in any one of claims 1-9, characterized in that, The system includes: The information acquisition module is used to acquire the current transaction execution information and current thermal status information of the transaction server, and to determine the execution time window of the key nodes of the transaction based on the current transaction execution information. The current thermal status information is used to determine the thermal disturbance status of the liquid cooling heat dissipation node. The timing conflict analysis module is used to determine the thermal disturbance time window for thermal disturbance propagation to the liquid cooling node corresponding to the key transaction node based on the preset transaction execution tree, heat flow propagation tree, and mapping relationship between transaction processing nodes and liquid cooling nodes in the transaction server. When the thermal disturbance time window and the execution time window meet the preset timing conflict conditions, the module reversely determines the source node of the thermal disturbance and the adjustable cooling path based on the heat flow propagation tree. The nodes of the transaction execution tree are used to represent transaction processing nodes, and the nodes of the heat flow propagation tree are used to represent liquid cooling nodes. The liquid cooling control module is used to generate a liquid cooling heat dissipation control strategy based on the heat disturbance source node and the adjustable cooling path, and to adjust the working state of the liquid cooling heat dissipation circuit of the transaction server according to the liquid cooling heat dissipation control strategy.

Images