An adaptive task partitioning pipeline optimization method and system

Abstract

The present application relates to the technical field of data analysis, in particular to a self-adaptive task segmentation pipeline optimization method and system, which establishes a heat conduction parameter set of each computing node through the physical topology of a liquid-cooled computing cluster; collects the operation data of the computing node and the cooling system in real time, calculates the performance trajectory of each computing node according to the heat conduction parameter set and the operation data; when it is determined according to the performance trajectory that a computing node will trigger frequency reduction, a new task segmentation decision is generated after the unexecuted task flow is intervened cooperatively; the pipeline execution is regulated according to the new task segmentation decision; the execution efficiency of the pipeline task in the liquid-cooled computing cluster can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to an adaptive task splitting pipeline optimization method and system. Background Technology

[0002] In today's rapidly evolving digital age, liquid-cooled computing clusters, with their efficient heat dissipation and powerful computing performance, have been widely adopted in numerous fields with extremely high computing resource demands, such as data centers and large-scale scientific computing. Liquid-cooled computing clusters ensure the stability of computing equipment under prolonged high-load operation by directly contacting or circulating coolant with the computing nodes to remove heat.

[0003] However, in actual operation, the computing nodes of a liquid-cooled computing cluster generate a significant amount of heat when executing tasks. Although the liquid cooling system can dissipate heat, the distribution and variation of heat are highly complex due to the different computational characteristics of different tasks. When computing nodes operate under high load for extended periods, or execute computationally intensive tasks with high heat output, their chip temperatures rise rapidly. Once the temperature exceeds the set frequency reduction trigger threshold, the computing node will automatically reduce its computing frequency to protect its hardware from damage, leading to a decrease in computing performance. Pipeline tasks typically consist of multiple interconnected task units executed sequentially. A decrease in the performance of one computing node can trigger a chain reaction, significantly reducing the execution efficiency of the entire pipeline, extending task completion time, and consequently affecting the overall system's response speed and processing capacity.

[0004] Therefore, an adaptive task splitting pipeline optimization method is needed to improve the execution efficiency of pipeline tasks in liquid-cooled computing clusters. Summary of the Invention

[0005] (1) Technical problems to be solved

[0006] The purpose of this invention is to provide an adaptive task splitting pipeline optimization method and system to solve the problem of reduced pipeline task execution efficiency in liquid-cooled computing clusters due to frequency reduction triggered by temperature-induced performance degradation of computing nodes.

[0007] (2) Technical solution

[0008] To achieve the above objectives, on the one hand, the present invention provides an adaptive task splitting pipeline optimization method, the method comprising:

[0009] S1. Obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connection relationship, computing node location, coolant flow rate and radiator capacity; establish the heat conduction parameter set of each computing node based on the physical topology.

[0010] S2. Real-time acquisition of operating data from computing nodes and the cooling system; extraction of current chip temperature and power consumption values ​​of each computing node, as well as real-time liquid temperature values ​​of key monitoring points in the cooling system; determination of the time delay constant of each computing node based on the heat conduction parameter set, and setting the length of the prediction time window based on the time delay constant; acquisition of the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue, and calculation of the heat power release curve of each task unit; superposition of the heat power release curves to generate the expected heat load input sequence of each computing node within the prediction time window; combining the expected heat load input sequence with the current chip temperature, current inlet liquid temperature, and time delay constant of each computing node, iterative calculation of the heat transfer equation step by step to obtain the temperature trajectory of each computing node; comparison of the temperature trajectory with the frequency reduction trigger temperature threshold of each computing node to obtain the degree of temperature exceedance, and inference of the available computing frequency at the corresponding moment based on the degree of temperature exceedance as the performance trajectory of each computing node.

[0011] S3. When it is determined from the performance trajectory that a computing node will trigger frequency reduction, a new task splitting decision is generated after collaborative intervention on the unexecuted task flow; the pipeline execution is adjusted according to the new task splitting decision.

[0012] Furthermore, the method for determining the time delay constant of each computation node based on the set of heat conduction parameters includes:

[0013] The heat capacity parameters of each computing node and the thermal resistance parameters between the computing node and the coolant are obtained from the heat conduction parameter set; the inherent thermal time constant of each computing node is calculated based on the heat capacity parameters and thermal resistance parameters.

[0014] Obtain the current inlet liquid temperature and coolant flow rate of each computing node from the operating data, calculate the convective heat transfer coefficient of the coolant based on the coolant flow rate, and determine the equivalent thermal resistance value under the current heat dissipation conditions in combination with the current inlet liquid temperature.

[0015] The dynamic thermal time constant of each computing node under the current operating conditions is calculated based on the equivalent thermal resistance and heat capacity parameters; the time delay constant of each computing node is obtained by weighted fusion of the dynamic thermal time constant and the inherent thermal time constant.

[0016] Furthermore, the method for superimposing the heat power release curves to generate the expected heat load input sequence for each computing node within the prediction time window includes:

[0017] Based on the pipeline task scheduling plan, obtain the start execution time and expected duration of each task unit on each computing node, align the thermal power release curve of each task unit to a unified time axis according to the start execution time, and accumulate the thermal power release curves of all task units to be executed on the same computing node moment by moment to obtain the local thermal load sequence of the computing node.

[0018] Based on the connection relationship of the liquid cooling loop, determine the set of upstream computing nodes for each computing node, calculate the contribution of the local heat load sequence of each computing node in the upstream computing node set to the temperature rise of the coolant, and then time-shift the temperature rise contribution according to the transmission delay of the coolant flowing from the upstream computing node to the current computing node and add it to the local heat load sequence of the current computing node to obtain the expected heat load input sequence of each computing node within the prediction time window.

[0019] Furthermore, the method for generating a new task splitting decision after collaboratively intervening in the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction includes:

[0020] Extract the algorithm type and batch size of each scheduled task unit from the unexecuted task flow. Determine the ratio of computational intensity to memory access intensity of the task unit based on the algorithm type. Determine the parallelism and continuous execution duration of the task unit based on the batch size. Combine the ratio, parallelism, and continuous execution duration with historical execution records to calculate the heat density index and heat persistence index of each task unit. Based on the heat density index and heat persistence index, classify the task units in the unexecuted task flow into high heat shock tasks and low heat shock tasks.

[0021] Based on the connection relationship between high thermal shock tasks and liquid cooling circuits, a spatiotemporal orchestration result is obtained through distributed orchestration. Based on the margin space between the predicted available computing frequency and the frequency reduction trigger temperature threshold in the performance trajectory of each computing node, the execution frequency of high thermal shock tasks is actively constrained and optimized to determine the execution frequency configuration that balances the task completion time and the thermal power release rate. The spatiotemporal orchestration result and the execution frequency configuration are integrated to generate a new task splitting decision.

[0022] Furthermore, the method for classifying task units in the unexecuted task flow into high-thermal-shock tasks and low-thermal-shock tasks based on heat density and thermal sustainability indices includes:

[0023] The heat density index and heat sustainability index of each task unit are used to construct a two-dimensional feature vector. The heat density index and heat sustainability index of historical tasks in the liquid-cooled computing cluster are collected as a clustering sample set. The clustering sample set is used as a similarity measure by the weighted Euclidean distance between the heat density index and the heat sustainability index. The clustering centers are iteratively updated until convergence, resulting in the first clustering center representing high heat load characteristics and the second clustering center representing low heat load characteristics.

[0024] Calculate the distance between the two-dimensional feature vector of each task unit to be scheduled and the first and second cluster centers, respectively. Compare the distance of the same task unit to the first cluster center with the distance to the second cluster center. Task units whose distance to the first cluster center is less than their distance to the second cluster center are classified as high thermal shock tasks, and task units whose distance to the second cluster center is less than their distance to the first cluster center are classified as low thermal shock tasks. Calculate the thermal load margin of each cooling branch based on the heat dissipation capacity of each cooling branch in the liquid cooling circuit connection relationship and the current liquid temperature state. For task units in the cluster boundary region in the classification results, perform a secondary judgment based on the thermal load margin of the cooling branch to which their target calculation node belongs. Boundary task units with insufficient thermal load margin are reclassified as high thermal shock tasks.

[0025] Furthermore, the method of using a weighted Euclidean distance between the heat density index and the heat persistence index as a similarity measure for the clustered sample set includes:

[0026] The number of series nodes and the coolant flow rate of each cooling branch are obtained based on the connection relationship of the liquid cooling circuit. Cooling branches whose ratio of the number of series nodes to the coolant flow rate is greater than a preset sensitivity threshold are marked as thermally coupled sensitive branches.

[0027] Thermal coupling sensitivity is obtained by calculating the proportion of computing nodes carried by thermally coupled sensitive branches to the total number of computing nodes in the liquid-cooled computing cluster. When the thermal coupling sensitivity is higher than the preset benchmark value, the weight coefficient of the thermal density index is increased and the weight coefficient of the thermal sustainability index is decreased. When the thermal coupling sensitivity is lower than the preset benchmark value, the weight coefficient of the thermal sustainability index is increased and the weight coefficient of the thermal density index is decreased. The weighted Euclidean distance of each two-dimensional feature vector in the clustered sample set is calculated using the adjusted weight coefficients as a similarity measure.

[0028] Furthermore, the method for obtaining spatiotemporal orchestration results by dispersing and orchestrating based on the connection relationship between high thermal shock tasks and liquid cooling loops includes:

[0029] Based on the connection relationship of the liquid cooling circuit, the topological position of each computing node on the coolant flow path is extracted. The heat transfer attenuation coefficient between any two computing nodes is calculated and a thermal coupling strength matrix is ​​constructed. High thermal shock tasks and available computing nodes are regarded as two types of vertices in a bipartite graph. The reciprocal of the elements in the thermal coupling strength matrix is ​​used as the edge weight to construct a task node allocation bipartite graph.

[0030] The task node allocation bipartite graph is used to solve the spatial allocation scheme by performing the maximum weight matching algorithm; the coolant delivery delay is determined according to the upstream and downstream positional relationship of the computing nodes allocated to each high thermal shock task; the execution time of tasks on the same cooling branch is staggered and arranged with the coolant delivery delay as a constraint; the spatial allocation scheme and the staggered arrangement together constitute the spatiotemporal arrangement result.

[0031] Furthermore, the method for calculating the heat transfer attenuation coefficient between any two computing nodes and constructing the thermal coupling strength matrix includes:

[0032] The coolant flow path between any two computing nodes is traversed according to the liquid cooling loop connection relationship. The physical length of each pipe segment on the path is accumulated to obtain the coolant flow distance. The transmission time of the coolant from the upstream computing node to the downstream computing node is calculated based on the coolant flow distance and the current coolant flow rate. The heat attenuation ratio of the coolant during the transmission process is calculated based on the transmission time and the heat dissipation coefficient per unit length of the pipe as the heat transfer attenuation coefficient. A matrix framework is constructed using each computing node in the liquid-cooled computing cluster as the row index and column index. The heat transfer attenuation coefficient between any two computing nodes is filled into the corresponding matrix position. For computing node pairs without a coolant flow path, their heat transfer attenuation coefficient is set to zero to obtain the thermal coupling strength matrix.

[0033] On the other hand, based on the same inventive concept, the present invention also provides an adaptive task splitting pipeline optimization system, the system comprising: a heat conduction parameter set establishment module, a performance trajectory calculation module, and a pipeline execution control module, wherein each module is sequentially connected in communication;

[0034] The heat conduction parameter set establishment module is used to obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connection relationship, computing node location, coolant flow rate and radiator capacity; and establishes the heat conduction parameter set of each computing node based on the physical topology.

[0035] The performance trajectory calculation module is used to collect real-time operating data of computing nodes and cooling systems, extract the current chip temperature and power consumption of each computing node, and the real-time liquid temperature of each key monitoring point of the cooling system from the operating data; determine the time delay constant of each computing node based on the heat conduction parameter set, and set the length of the prediction time window based on the time delay constant; obtain the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue, and calculate the heat power release curve of each task unit; superimpose the heat power release curves to generate the expected heat load input sequence of each computing node within the prediction time window; combine the expected heat load input sequence with the current chip temperature, current inlet liquid temperature and time delay constant of each computing node to calculate the temperature trajectory of each computing node step by step through the heat transfer equation; compare the temperature trajectory with the frequency reduction trigger temperature threshold of each computing node to obtain the degree of temperature exceedance, and infer the available computing frequency at the corresponding time as the performance trajectory of each computing node based on the degree of temperature exceedance.

[0036] The pipeline execution control module is used to generate a new task splitting decision after coordinating intervention on the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction; and to control pipeline execution based on the new task splitting decision.

[0037] (3) Beneficial effects

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. By acquiring the physical topology of the liquid-cooled computing cluster and establishing a set of heat conduction parameters, combined with real-time collected operating data, the performance trajectory of each computing node can be accurately calculated. This enables proactive prediction of frequency reduction risks, effectively avoiding pipeline task interruptions or sudden drops in execution efficiency caused by sudden frequency reduction of computing nodes, and providing a basic guarantee for the stable execution of pipeline tasks.

[0040] 2. When it is determined that a computing node will trigger frequency reduction, the system will coordinately intervene in the unexecuted task flow and generate new task splitting decisions. It fully considers the characteristics of task units, classifies task units by calculating heat density and thermal sustainability indicators, and performs reasonable spatiotemporal orchestration and execution frequency configuration based on the liquid cooling loop connection relationship and computing node performance trajectory. It can dynamically adjust task allocation according to actual conditions, making the distribution of tasks in the cluster more reasonable, and further improving the execution efficiency and stability of pipeline tasks when facing potential frequency reduction risks.

[0041] 3. By rationally distributing and staggering the execution of high thermal shock tasks, and by actively constraining and optimizing the execution frequency of tasks, the optimal allocation of computing and cooling resources within the cluster is achieved. This enables the entire liquid-cooled computing cluster to operate more efficiently when facing complex and variable task loads, maximizing the overall performance of the cluster and providing strong support for the efficient completion of large-scale computing tasks. Attached Figure Description

[0042] Figure 1 This is a flowchart of an adaptive task splitting pipeline optimization method according to Embodiment 1 of the present invention.

[0043] Figure 2 This is a schematic diagram of the module composition of an adaptive task splitting pipeline optimization system according to Embodiment 2 of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: As Figure 1 As shown, this embodiment provides an adaptive task splitting pipeline optimization method, the method including:

[0046] S1. Obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connections, computing node locations, coolant flow rate, and radiator capacity. Establish a set of heat conduction parameters for each computing node based on the physical topology. During the initialization phase, the system reads the configuration information of the liquid-cooled computing cluster. This configuration information describes the physical layout of all computing nodes within the cluster and the flow path of the coolant. The liquid cooling loop connections record which computing nodes are located on the same cooling branch and the order of these nodes in the coolant flow direction. The computing node locations further indicate the physical coordinates of each node within the rack. The coolant flow rate refers to the volumetric velocity of the coolant in each cooling branch, directly affecting the coolant's ability to remove heat. The radiator capacity describes the heat exchange efficiency of the radiators configured for each computing node; different radiator specifications have different heat dissipation capacities under the same liquid temperature conditions. The set of heat conduction parameters contains key parameters describing the thermodynamic characteristics of the computing nodes, mainly covering heat capacity parameters, thermal resistance parameters, and the thermal coupling coefficient with adjacent computing nodes. The heat capacity parameter reflects the magnitude of temperature change when a computing node absorbs or releases a unit of heat; its value is related to the chip material, packaging structure, and heat sink quality. The thermal resistance parameter characterizes the resistance encountered during heat transfer from the chip to the coolant, including the interfacial thermal resistance between the chip and the heat sink, the internal thermal conductivity of the heat sink, and the convective thermal resistance between the heat sink and the coolant. The thermal coupling coefficient quantifies the impact of an upstream computing node on the heat dissipation of a downstream computing node; this coefficient is related to the pipe length between the two nodes, the coolant flow rate, and the insulation performance of the pipe.

[0047] S2. Real-time acquisition of operating data from computing nodes and cooling systems; The system continuously reads current chip temperature and real-time power consumption data from sensors on each computing node, and at the same time obtains liquid temperature values ​​at key locations from monitoring points of the cooling system as operating data. These operating data, together with the aforementioned set of heat conduction parameters, constitute the input for performance trajectory calculation.

[0048] The system extracts the current chip temperature and power consumption values ​​of each computing node, as well as the real-time liquid temperature values ​​of key monitoring points in the cooling system, from the operational data. It determines the time delay constant of each computing node based on the heat conduction parameter set and sets the length of the prediction time window based on this time delay constant. The system reads the current chip temperature and power consumption values ​​from the onboard sensors of each computing node. The chip temperature is typically measured by temperature sensors embedded inside the GPU or CPU, while the power consumption value is obtained through real-time monitoring by the power management unit. The system collects real-time liquid temperature values ​​from key monitoring points in the cooling system, including the outlet location of the cooling unit, the inlet location of each cooling branch, and temperature probes at the contact points between each computing node and the coolant.

[0049] After acquiring the real-time status, the system needs to determine the length of the prediction time window, i.e., the time range over which temperature changes need to be predicted. The prediction time window is set based on the time delay constant of each computing node. The time delay constant is a physical quantity describing the response speed of a thermal system; it represents the time required for the temperature of a computing node to change from its initial value to its final steady-state value when a step change occurs in the heat input. A larger time delay constant indicates greater thermal inertia of the computing node and a slower temperature change; conversely, a smaller time delay constant indicates a faster temperature response. The prediction time window length is typically set to 3 to 5 times the maximum time delay constant to ensure that the prediction range covers the entire temperature change process.

[0050] The system acquires the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue, and calculates the heat release curve of each task unit. These heat release curves are then superimposed to generate the expected heat load input sequence for each computing node within the prediction time window. The expected heat load input sequence describes the change in heat power input that each node will experience over a future period. Information about each task unit is extracted from the current in-transit task queue, which contains all scheduled but not yet completed tasks. The estimated computational load is determined based on the task's algorithm type and input data size. For example, the computational load of a convolutional neural network forward propagation task with a batch size of 64 can be estimated by the number of floating-point operations in the network structure. Historical power consumption characteristics are derived from the power consumption records of the same or similar tasks executed in the past. The system maintains a task power consumption feature library, recording typical power consumption curves for various tasks at different execution frequencies. The heat release curve is a curve with time on the horizontal axis and heat power on the vertical axis, describing the rate of heat release from the start to the end of the task. Different types of tasks have different heat power release curves: compute-intensive tasks release high heat power continuously throughout the execution period, and their curves are plateau-shaped; memory-intensive tasks consume less power during the data loading phase but more power during the computation phase, and their curves are pulse-shaped; hybrid tasks exhibit periodic fluctuations.

[0051] The system iteratively calculates the temperature trajectory of each computing node by combining the expected thermal load input sequence with the current chip temperature, current inlet liquid temperature, and time delay constant of each computing node through the heat transfer equation, step by step. The temperature trajectory is then compared with the frequency reduction trigger temperature threshold of each computing node to determine the degree of temperature exceedance. Based on the degree of temperature exceedance, the available computing frequency at the corresponding moment is inferred as the performance trajectory of each computing node. The heat transfer equation describes the physical law of the computing node temperature change over time. Its basic form is: the rate of change of node temperature equals the thermal input power minus the heat dissipation power divided by the heat capacity. The heat dissipation power is proportional to the temperature difference between the current node temperature and the inlet liquid temperature, and the reciprocal of the proportionality coefficient is the thermal resistance. The system uses the current chip temperature as the initial condition, the expected thermal load input sequence as the driving input, and the current inlet liquid temperature as the boundary condition. It iteratively solves the heat transfer equation step by step using numerical integration methods such as the Euler method or the Runge-Kutta method. The step size of each time step is typically 1 second. The system calculates the node temperature sequentially for the next second, the next second after that, until the end of the prediction time window, thus obtaining the complete temperature trajectory.

[0052] The frequency reduction trigger temperature threshold is a manufacturer-preset safe temperature upper limit. Different GPUs or CPUs have different threshold settings, with common values ​​between 80 and 95 degrees Celsius. The system checks the temperature value in the temperature trajectory every moment to see if it exceeds the frequency reduction trigger temperature threshold. If it does, the extent of the exceedance is recorded as the degree of temperature limit violation. The available computing frequency at the corresponding moment is inferred based on the degree of temperature limit violation. The chip's frequency reduction mechanism usually follows a graded strategy: a small frequency reduction when the temperature slightly exceeds the threshold, and a deep frequency reduction when the temperature significantly exceeds the threshold. The system maps the degree of temperature limit violation to the frequency reduction ratio based on the chip's frequency reduction characteristic curve, thereby obtaining the available computing frequency at each moment. By concatenating the available computing frequencies at all moments, the performance trajectory of the computing node is formed. For example, the frequency reduction trigger temperature threshold of a computing node is 85 degrees Celsius, the current chip temperature is 72 degrees Celsius, and the prediction time window is 120 seconds. Through iterative calculations using the heat transfer equation, the system obtained the temperature trajectory of this node: The temperature slowly rises to 78 degrees Celsius for the first 40 seconds; from the 40th to the 70th second, due to the execution of a high-power task, the temperature rapidly climbs to 88 degrees Celsius; from the 70th to the 90th second, the temperature remains between 87 and 89 degrees Celsius; after the 90th second, the temperature gradually drops as the task completes. Comparing this temperature trajectory with the 85-degree Celsius threshold, the temperature begins to exceed the limit from the 45th second, reaching a maximum exceedance of 4 degrees Celsius. Based on the chip's frequency reduction characteristics, the frequency decreases by 5% for every 1-degree Celsius increase in temperature above the threshold. Therefore, from the 45th to the 90th second, the available computing frequency drops to 80% to 85% of the base frequency; after the 90th second, the frequency gradually recovers as the temperature falls back. This curve showing the change in available computing frequency over time represents the performance trajectory of this node.

[0053] S3. When the performance trajectory indicates that a computing node will trigger frequency reduction, a new task splitting decision is generated after collaborative intervention on the unexecuted task flow. The pipeline execution is then adjusted based on the new task splitting decision. The adjustment operations include sending updated task allocation instructions to the task scheduler, issuing execution frequency configuration parameters to each computing node, and adjusting the dependency waiting time between tasks. After adjustment, the pipeline will run according to the optimized scheme, and the temperature of each computing node will be controlled within a safe range, avoiding a sudden drop in performance due to frequency reduction, thereby ensuring that the entire pipeline tasks can be executed in a stable and efficient manner.

[0054] The method for determining the time delay constant of each computing node based on the set of heat conduction parameters includes:

[0055] The thermal capacity parameters and thermal resistance parameters between the computing nodes and the coolant are obtained from the heat conduction parameter set for each computing node. The inherent thermal time constant of each computing node is calculated based on these parameters. The physical meaning of the thermal capacity parameter is the amount of heat absorbed to raise the temperature of a computing node by 1 degree Celsius, measured in joules per degree Celsius. The value of the thermal capacity parameter depends on the chip's mass, the material's specific heat capacity, and the mass and material of the heat sink. For example, the thermal capacity of a data center GPU chip is approximately 50 to 100 joules per degree Celsius, and with the heat sink, the overall thermal capacity can reach 200 to 500 joules per degree Celsius. The physical meaning of the thermal resistance parameter is the temperature difference corresponding to a unit heat flux during heat transfer from the chip to the coolant, measured in degrees Celsius per watt. A larger thermal resistance parameter indicates greater resistance to the heat dissipation path, making it more difficult for heat to transfer. The inherent thermal time constant is equal to the product of the thermal capacity parameter and the thermal resistance parameter, and its physical meaning describes the node's thermal response speed under standard operating conditions.

[0056] The system obtains the current inlet liquid temperature and coolant flow rate of each computing node from the operational data. Based on the coolant flow rate, it calculates the convective heat transfer coefficient of the coolant and, combined with the current inlet liquid temperature, determines the equivalent thermal resistance under the current heat dissipation conditions. The inlet liquid temperature affects the driving temperature difference for heat dissipation; the lower the liquid temperature, the larger the temperature difference and the higher the heat dissipation efficiency. The coolant flow rate affects the convective heat transfer efficiency; the larger the flow rate, the stronger the coolant's ability to remove heat. The convective heat transfer coefficient describes the intensity of heat exchange between the coolant and the radiator surface, and its value is related to the coolant's physical properties, flow state, and the radiator's surface structure. Under turbulent conditions, the convective heat transfer coefficient is proportional to the coolant flow velocity raised to the power of 0.8. The system uses empirical formulas or lookup tables to determine the convective heat transfer coefficient under the current flow rate conditions, and then calculates the equivalent thermal resistance under the current heat dissipation conditions based on the inlet liquid temperature. The equivalent thermal resistance reflects the actual resistance to heat transfer from the chip to the coolant under the current actual operating conditions. Its value is usually smaller than the thermal resistance parameter under standard operating conditions because the cooling system is often in a more optimal heat dissipation state during actual operation.

[0057] The dynamic thermal time constant of each computing node under the current operating conditions is calculated based on the equivalent thermal resistance and heat capacity parameters. This dynamic thermal time constant is then weighted and fused with the intrinsic thermal time constant to obtain the time delay constant of each computing node. Multiplying the equivalent thermal resistance by the heat capacity parameter yields the dynamic thermal time constant of each computing node under the current operating conditions. The dynamic thermal time constant reflects the thermal response speed of the node under actual heat dissipation conditions and describes the rate of temperature change under the current state more accurately than the intrinsic thermal time constant. The purpose of weighted fusion is to balance the intrinsic thermal characteristics of the node with the current operating state. The fusion formula adopts a linear weighting form, and the weight coefficients are determined according to the stability of the cooling system: when the cooling system is in steady-state operation, the dynamic thermal time constant has a larger weight; when the cooling system is in a transitional state, the intrinsic thermal time constant has a larger weight. For example, the weight configuration is 70% for the dynamic thermal time constant and 30% for the intrinsic thermal time constant. The time delay constant obtained after weighted fusion reflects both the intrinsic thermal characteristics of the node and adapts to the current actual operating conditions.

[0058] The method for superimposing the heat power release curves to generate the expected heat load input sequence for each computing node within the prediction time window includes:

[0059] Based on the pipeline task scheduling plan, the system obtains the start execution time and estimated duration of each task unit on each computing node, and aligns the thermal power release curves of each task unit to a unified time axis according to their start execution times. The system then accumulates the thermal power release curves of all task units to be executed on the same computing node moment by moment to obtain the local thermal load sequence of the computing node. The system also obtains the execution schedule of each task unit from the pipeline task scheduling plan, including which computing node each task unit is assigned to, its start execution time on that node, and its estimated duration. The start execution time is determined by the pipeline dependencies and the queuing order of the task queue, while the estimated duration is estimated based on the computational load of the task and the processing capacity of the target node. After obtaining this scheduling information, the system aligns the thermal power release curves of each task unit to a unified time axis according to their start execution times. For example, if task A starts execution at 10 seconds and lasts for 20 seconds, and task B starts execution at 25 seconds and lasts for 15 seconds, then the thermal power release curve of task A is placed in the interval from 10 seconds to 30 seconds on the time axis, and the thermal power release curve of task B is placed in the interval from 25 seconds to 40 seconds.

[0060] For multiple task units assigned to the same compute node, the system accumulates their heat release curves time-by-time. The accumulation process follows the principle of heat power superposition, meaning that the total heat power borne by the node at any given time is equal to the sum of the heat power of all tasks currently executing. During pipelined execution, due to dependencies between tasks, tasks on the same node are usually executed sequentially rather than in parallel; therefore, most of the time only one task is releasing heat power. However, during task switching transitions, the tail heat power of the previous task may overlap with the starting heat power of the next task. In this case, accumulation is necessary. After accumulation, a local heat load sequence for each compute node is obtained, which describes the change in heat power generated by the compute node's own running tasks over time.

[0061] Based on the liquid cooling loop connection relationship, the upstream computing node set of each computing node is determined. The contribution of the local heat load sequence of each computing node in the upstream computing node set to the temperature rise of the coolant is calculated. This temperature rise contribution is then time-shifted according to the transmission delay of the coolant flowing from the upstream computing node to the current computing node and added to the local heat load sequence of the current computing node to obtain the expected heat load input sequence of each computing node within the prediction time window. In a series liquid cooling system, the coolant flows through each computing node sequentially. The heat released by the upstream node will raise the coolant temperature. When this heated coolant flows to the downstream node, it will cause the inlet liquid temperature of the downstream node to rise, resulting in a decrease in heat dissipation efficiency. This heat transfer relationship between upstream and downstream nodes is called the thermal coupling effect, which is an important feature that distinguishes liquid-cooled computing clusters from air-cooled systems. To accurately describe the thermal coupling effect, the system determines the upstream computing node set of each computing node based on the liquid cooling loop connection relationship. Upstream computing nodes refer to all nodes located before the current node on the coolant flow path. The calculation of the temperature rise contribution is based on the principle of energy conservation: the heat released by the upstream node is absorbed by the coolant, causing the coolant temperature to rise. The temperature rise is equal to the absorbed heat divided by the product of the coolant's mass flow rate and specific heat capacity.

[0062] It takes time for coolant to flow from the upstream node to the current node. During this time, the coolant temperature will decrease due to heat dissipation through the pipes, but some heat will still be transferred downstream. The system calculates the transmission delay time based on the pipe length and coolant flow rate between the upstream and downstream nodes, and then times-shifts the temperature rise contribution according to this transmission delay. Time-shifting means delaying the temperature rise effect generated by the upstream node at a certain moment before it acts on the downstream node. The expected heat load input sequence includes not only the heat power generated by the node's own tasks but also the equivalent increase in heat load due to heat transfer from the upstream node.

[0063] The method for generating a new task splitting decision after collaboratively intervening in the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction includes:

[0064] The algorithm type and batch size of each scheduled task unit are extracted from the unexecuted task flow. Based on the algorithm type, the ratio of computational intensity to memory access intensity of the task unit is determined. Based on the batch size, the parallelism and execution duration of the task unit are determined. These ratios, parallelism, and execution duration are combined with historical execution records to calculate the heat density and heat persistence indices of each task unit. Based on these indices, the task units in the unexecuted task flow are divided into high-heat-impact tasks and low-heat-impact tasks. The first step of collaborative intervention is to perform thermal characteristic analysis and classification on each scheduled task unit in the unexecuted task flow. The system extracts two key attributes from each scheduled task unit: algorithm type and batch size. The algorithm type describes the computational logic executed by the task; different algorithm types have different ratios of computational intensity to memory access intensity. Computationally intensive algorithms mainly use floating-point operations, continuously occupying computing units and generating a large amount of heat; memory-intensive algorithms mainly use data reading and writing, with lower utilization of computing units and relatively less heat generation. Taking deep learning as an example, convolutional layers and fully connected layers are computationally intensive, while data preprocessing and feature concatenation operations are memory-intensive. The system determines the ratio of computational intensity to memory intensity of a task unit based on the algorithm type. This ratio directly affects the power consumption and heat release intensity during task execution. Batch size describes the amount of data processed by a task unit at one time, determining the parallelism and execution duration of the task. The larger the batch size, the higher the parallelism of the task, the higher the utilization rate of the computing unit, and the longer the execution duration of the task. Batch size and algorithm type together determine the overall thermal load characteristics of the task. The heat density index describes the intensity of heat released by a task unit per unit time; a higher value indicates a greater instantaneous heat power. The heat density index is calculated by dividing the peak heat power of the task by its computational load and then multiplying by a correction factor determined according to the algorithm type. The thermal persistence index describes the duration for which a task unit continuously releases high heat; a higher value indicates a longer period of continuous pressure on the cooling system. The thermal sustainability metric is calculated by multiplying the percentage of time during task execution when the thermal power exceeds a certain baseline value by the total execution time of the task. High thermal shock tasks are those with high heat density or thermal sustainability metrics. These tasks place significant stress on the cooling system and are a major factor leading to compute node temperatures exceeding limits. Low thermal shock tasks are those with both metrics being relatively low, and these tasks have a smaller impact on the cooling system.

[0065] Based on the connection relationship between high thermal shock tasks and liquid cooling circuits, a distributed orchestration is performed to obtain spatiotemporal orchestration results. Based on the margin between the predicted available computing frequency and the frequency reduction trigger temperature threshold in the performance trajectory of each computing node, active constraint optimization is performed on the execution frequency of high thermal shock tasks to determine an execution frequency configuration that balances task completion time and heat power release rate. The spatiotemporal orchestration results and execution frequency configuration are then integrated to generate a new task splitting decision. The purpose of distributed orchestration is to avoid multiple high thermal shock tasks being concentrated on the same cooling branch or executed within similar time periods, thereby preventing excessive heat accumulation in local areas. Execution frequency directly affects the power consumption and heat power release rate of a task: higher frequencies result in faster computation but more heat generation, while lower frequencies result in less heat generation but slower computation. The goal of active constraint optimization is to find an execution frequency configuration that balances task completion time and heat power release rate within the allowable range of the performance trajectories of each computing node. The system extracts the margin space between the available computing frequency and the frequency reduction trigger temperature threshold at each moment from the performance trajectory. The margin space represents how much additional thermal load a node can withstand without triggering passive frequency reduction. Based on the margin space, an optimization problem is constructed: finding the execution frequency configuration that minimizes the total peak thermal power while satisfying the margin space constraint and the task deadline constraint. This optimization problem can be solved using a greedy algorithm or linear programming method to obtain the optimal execution frequency for each high thermal shock task. The new task partitioning decision fully describes how each scheduled task unit should be executed: which computing node to allocate to, when to start execution, and at what frequency. The system distributes this decision to the task scheduler and the frequency management module of each computing node to control the pipeline execution.

[0066] The method for classifying task units in an unexecuted task flow into high-thermal-shock tasks and low-thermal-shock tasks based on heat density and thermal sustainability indices includes:

[0067] A two-dimensional feature vector is constructed using the heat density and thermal sustainability indices of each task unit. The heat density and thermal sustainability indices of historically executed tasks in the liquid-cooled computing cluster are collected as a clustering sample set. The clustering sample set is then iteratively updated with the weighted Euclidean distance between the heat density and thermal sustainability indices as a similarity metric until convergence, resulting in a first cluster center representing high heat load characteristics and a second cluster center representing low heat load characteristics. The heat density and thermal sustainability indices of each task unit are then used to construct a two-dimensional feature vector, with each task unit corresponding to a point in the feature space. The horizontal axis represents the heat density value, and the vertical axis represents the thermal sustainability value. This two-dimensional representation intuitively depicts the thermal characteristics of the task: points located in the upper right corner of the feature space represent tasks with high heat density and long-term continuous heat release, while points located in the lower left corner represent tasks with low heat density and short-term heat release.

[0068] To establish classification criteria, the system collected heat density and thermal sustainability indices from historically executed tasks in the liquid-cooled computing cluster as a clustering sample set. The historical data originated from task execution records accumulated during the system's long-term operation; each record contained task type information and its actual measured heat density and thermal sustainability values. The clustering algorithm started with two randomly initialized cluster centers and iteratively executed the following steps: assigning each sample to the cluster belonging to the nearest cluster center, and then recalculating the mean of all samples within each cluster as the new cluster center. This iteration continued until the cluster centers no longer changed significantly, at which point the algorithm converged. After clustering, the system obtained two cluster centers. The cluster center located in the region with higher heat density and thermal sustainability in the feature space was labeled as the first cluster center, representing high heat load characteristics; the cluster center located in the lower region was labeled as the second cluster center, representing low heat load characteristics. The first and second cluster centers constituted the classification reference.

[0069] The system calculates the distances between the two-dimensional feature vectors of each task unit to be scheduled and the first and second cluster centers. It compares the distances from the same task unit to the first and second cluster centers, classifying task units whose distances to the first and second cluster centers are less than their distances to the first and second cluster centers as high-thermal-shock tasks, and those whose distances are less than their distances to the second and first cluster centers as low-thermal-shock tasks. Based on the heat dissipation capacity of each cooling branch in the liquid cooling circuit connection relationship and the current liquid temperature, the system calculates the thermal load margin of each cooling branch. For task units located in the cluster boundary region in the classification results, a secondary judgment is made based on the thermal load margin of the cooling branch to which their target calculation node belongs. Boundary task units with insufficient thermal load margins are reclassified as high-thermal-shock tasks. The distance calculation uses the weighted Euclidean distance formula. The system compares the distances of the same task unit to the two cluster centers: if the distance to the first cluster center is less than the distance to the second cluster center, the task unit is classified as a high-thermal-shock task; if the distance to the second cluster center is even smaller, it is classified as a low-thermal-shock task.

[0070] After initial classification, the system performs a secondary assessment of task units located in the cluster boundary region. The cluster boundary region refers to an area in the feature space where two cluster centers are close together; the classification result for task units located in this region is subject to some uncertainty. The secondary assessment comprehensively considers the actual thermal load of the cooling branch to which the target computing node belongs. The system calculates the thermal load margin of each branch based on the heat dissipation capacity of each cooling branch in the liquid cooling circuit connection relationship and the current liquid temperature. The heat dissipation capacity is determined by the coolant flow rate and radiator specifications of the cooling branch, while the current liquid temperature is obtained in real-time from the operating data. The thermal load margin indicates how much additional heat load the branch can withstand under the current condition without causing any node temperature to exceed its limit. For task units initially classified as low thermal shock type but located in the boundary region, the system checks the thermal load margin of the cooling branch to which its target computing node belongs. If the thermal load margin of a cooling branch is insufficient—for example, if the thermal load margin is less than the system's preset safety margin threshold—it means that even tasks with moderate thermal characteristics may pose an overheating risk to that branch. In this case, the system reclassifies the boundary task unit as a high-thermal-shock task to ensure that it is handled more cautiously in subsequent orchestration. This secondary judgment mechanism allows task classification to adapt to the real-time operating status of the cluster, improving the accuracy and practicality of the classification results.

[0071] The method of using a weighted Euclidean distance between heat density and heat persistence indices as a similarity measure for clustered sample sets includes:

[0072] The number of series nodes and coolant flow rate of each cooling branch are obtained based on the connection relationship of the liquid cooling loop. Cooling branches whose ratio of the number of series nodes to the coolant flow rate is greater than a preset sensitivity threshold are marked as thermally coupled sensitive branches. The setting of the weight coefficients directly affects the sensitivity of the clustering algorithm to the thermal characteristics of the task. If the weight of the heat density index is large, the algorithm pays more attention to the instantaneous heat power intensity of the task and tends to identify high-power tasks as high-thermal-shock tasks. If the weight of the thermal sustainability index is large, the algorithm pays more attention to the continuous heat release time of the task and tends to identify long-running tasks as high-thermal-shock tasks. The reasonable configuration of the weights should match the actual heat dissipation characteristics of the liquid-cooled computing cluster. The basis for weight adjustment is the thermal coupling sensitivity of the liquid cooling loop. In clusters with high thermal coupling sensitivity, the heat from upstream nodes has a more significant impact on downstream nodes. In such clusters, tasks with instantaneous high thermal power pose a greater risk because the resulting heat peaks are transmitted along the coolant flow direction and accumulate at downstream nodes. In clusters with low thermal coupling sensitivity, the thermal effects between nodes are relatively independent. In these clusters, tasks with prolonged continuous heat release pose a greater risk because the continuous heat load causes the node temperature to gradually increase. To quantify thermal coupling sensitivity, the system first obtains the structural parameters of each cooling branch based on the liquid cooling loop connection relationship, including the number of series nodes and the coolant flow rate. The number of series nodes refers to the number of computing nodes connected sequentially on a cooling branch; a larger number indicates a longer thermal coupling chain, and the more pronounced the cumulative effect of upstream influence on downstream nodes. The coolant flow rate refers to the speed at which the coolant flows in the branch; a faster flow rate results in shorter coolant transmission time between nodes, less heat attenuation, and a more significant thermal coupling effect. The system calculates the ratio of the number of series nodes in each cooling branch to the coolant flow rate. This ratio reflects the thermal coupling sensitivity of the branch: the larger the ratio, the more nodes and the slower the flow rate, and the weaker the thermal coupling effect; the smaller the ratio, the fewer nodes and the faster the flow rate, or the more nodes and the faster the flow rate, and the stronger the thermal coupling effect.

[0073] Thermal coupling sensitivity is calculated by proportionally analyzing the number of computing nodes on thermally sensitive branches relative to the total number of computing nodes in the liquid-cooled computing cluster. When the thermal coupling sensitivity exceeds a preset benchmark, the weighting coefficient of the heat density index is increased while the weighting coefficient of the heat sustainability index is decreased. Conversely, when the thermal coupling sensitivity falls below the preset benchmark, the weighting coefficient of the heat sustainability index is increased while the weighting coefficient of the heat density index is decreased. The adjusted weighting coefficients are used to calculate a weighted Euclidean distance on each two-dimensional feature vector in the clustered sample set as a similarity measure. For example, if the liquid-cooled cluster has 64 computing nodes, with 40 nodes located on thermally sensitive branches, the thermal coupling sensitivity is 62.5%. The system compares the thermal coupling sensitivity with the preset benchmark and adjusts the weighting coefficients accordingly. When the thermal coupling sensitivity exceeds the preset benchmark, it indicates that most nodes in the cluster are located on thermally sensitive cooling branches, posing a significant risk of impact on downstream nodes from instantaneous high-power tasks. In this case, the system increases the weighting coefficient of the heat density index and correspondingly decreases the weighting coefficient of the heat sustainability index, making the clustering algorithm more sensitive to identifying high-power tasks. When the thermal coupling sensitivity is lower than the preset benchmark value, it indicates that the thermal coupling effect of most nodes in the cluster is weak, and the risk of temperature accumulation caused by long-term continuous heat release is more worthy of attention. At this time, the system increases the weight coefficient of the thermal persistence index and decreases the weight coefficient of the heat density index, so that the clustering algorithm pays more attention to long-running tasks.

[0074] The magnitude of the weight adjustment is proportional to the degree to which the thermal coupling sensitivity deviates from the baseline value. For example, with a preset baseline value of 50%, both the heat density weight and the thermal sustainability weight are 0.5 under the baseline condition. If the actual thermal coupling sensitivity is 70%, exceeding the baseline by 20 percentage points, the system will increase the heat density weight to 0.6 and decrease the thermal sustainability weight to 0.4; if the actual thermal coupling sensitivity is 30%, below the baseline by 20 percentage points, the heat density weight will decrease to 0.4 and the thermal sustainability weight will increase to 0.6. Using the adjusted weight coefficients, the system calculates the weighted Euclidean distance for each two-dimensional feature vector in the clustered sample set and uses this as a similarity measure for the clustering algorithm. This adaptive weight adjustment mechanism enables the task classification criteria to match the actual heat dissipation architecture of the liquid-cooled computing cluster, accurately identifying the task types that pose the greatest threat to system stability on clusters with different configurations.

[0075] The method for obtaining spatiotemporal orchestration results by dispersing and orchestrating based on the connection relationship between high thermal shock tasks and liquid cooling circuits includes:

[0076] Based on the liquid cooling loop connection relationship, the topological position of each computing node on the coolant flow path is extracted. The heat transfer attenuation coefficient between any two computing nodes is calculated, and a thermal coupling strength matrix is ​​constructed. High thermal shock tasks and available computing nodes are treated as two types of vertices in a bipartite graph. The reciprocal of the elements in the thermal coupling strength matrix is ​​used as the edge weight to construct a task node allocation bipartite graph. The goal of distributed orchestration is to rationally allocate high thermal shock tasks to each computing node and schedule appropriate execution times to ensure a balanced distribution of heat load across the branches of the liquid cooling loop and avoid local overheating. The orchestration process consists of two stages: spatial allocation and temporal orchestration. The spatial allocation stage requires quantifying the thermal coupling relationship between computing nodes. The system extracts the topological position of each computing node on the coolant flow path based on the liquid cooling loop connection relationship. The topological position describes which cooling branch the node is located on and its sequential position on that branch. A bipartite graph is a special graph structure where the vertices are divided into two disjoint sets, and edges exist only between vertices in different sets. In the task node allocation bipartite graph, one type of vertices corresponds to high thermal shock tasks, and the other type corresponds to available computing nodes. If a task can be assigned to a specific computing node, an edge is drawn between the corresponding task vertex and the node vertex. The system assigns nodes with strong thermal coupling to different tasks to avoid multiple thermally impactful tasks running simultaneously on nodes with significant mutual influence. To this end, the system uses the reciprocal of the elements in the thermal coupling strength matrix as the edge weights. If the heat transfer attenuation coefficient between two nodes is large, their thermal coupling is strong, and the corresponding edge weight is small; if the heat transfer attenuation coefficient is small, the thermal coupling is weak, and the edge weight is large. By setting these weights, selecting the set of edges with the largest total weight is equivalent to selecting the task node allocation scheme with the weakest thermal coupling.

[0077] The system assigns a maximum weight matching algorithm to the bipartite graph of task nodes to solve for the spatial allocation scheme. Based on the upstream and downstream positional relationships of the computing nodes assigned to each high thermal shock task, the coolant transfer delay is determined. Using this coolant transfer delay as a constraint, the execution times of tasks on the same cooling branch are staggered. The spatial allocation scheme and staggered scheduling together constitute the spatiotemporal arrangement result. Maximum weight matching is a classic problem in combinatorial optimization, aiming to find a set of edges in a bipartite graph such that each vertex is covered by at most one edge, and the sum of the weights of the selected edges is maximized. Commonly used algorithms include the Hungarian algorithm and the Kuhn-Munkres algorithm, which can obtain the optimal solution in polynomial time. The matching result output by the algorithm is the spatial allocation scheme, indicating which computing node each high thermal shock task should be assigned to for execution. Due to the thermal coupling effect of the liquid cooling circuit, there is a transmission delay; heat generated by upstream nodes takes a certain amount of time to affect downstream nodes. The system determines the coolant transfer delay based on the upstream and downstream positional relationships of the computing nodes assigned to each high thermal shock task. The transfer delay is equal to the pipe length between upstream and downstream nodes divided by the coolant flow rate. The principle of staggered scheduling is as follows: if task A is assigned to an upstream node and task B to a downstream node, the start time of task B's execution should be staggered from the peak thermal power time of task A by at least one transmission delay time interval. This ensures that when the heat generated by task A is transferred to the node where task B is located, task B has not yet entered its peak thermal power stage. This arrangement can prevent the heat transferred from the upstream from superimposing with the heat generated by the downstream task itself, resulting in a higher temperature peak. The spatiotemporal orchestration result assigns an execution node and execution time to each high thermal shock task, achieving spatial dispersion and temporal staggering of the thermal load, minimizing the risk of computing node frequency reduction due to heat accumulation.

[0078] The method for calculating the heat transfer attenuation coefficient between any two computing nodes and constructing the thermal coupling strength matrix includes:

[0079] The system traverses the coolant flow path between any two computing nodes based on the liquid cooling loop connection relationship, accumulating the physical length of each pipe segment along the path to obtain the coolant flow distance. It then calculates the coolant transmission time from the upstream computing node to the downstream computing node based on the coolant flow distance and the current coolant flow rate. Finally, it calculates the heat attenuation ratio of the coolant during transmission based on the transmission time and the heat dissipation coefficient per unit length of the pipe, using this as the heat transfer attenuation coefficient. A matrix framework is constructed using each computing node in the liquid-cooled computing cluster as row and column indices. The heat transfer attenuation coefficient between any two computing nodes is filled into the corresponding matrix positions. For computing node pairs without a coolant flow path, their heat transfer attenuation coefficient is set to zero, resulting in a thermal coupling strength matrix. The system traverses the coolant flow path between any two computing nodes based on the liquid cooling loop connection relationship, identifying all pipe segments traversed by the coolant from the source node to the target node. The traversal process follows the actual flow direction of the coolant. If two nodes are located in the same cooling branch, the path is a sequence of pipe segments connecting them; if two nodes are located in different cooling branches, there is no direct coolant flow path. After determining the flow path, the system accumulates the physical lengths of each pipe segment along that path to obtain the coolant flow distance. The physical length information is derived from the design drawings or on-site measurement data of the liquid-cooled computing cluster, recording the actual length of each pipe segment. Based on the coolant flow distance and the current coolant flow velocity, the system calculates the transmission time of the coolant from the upstream computing node to the downstream computing node. The transmission time equals the flow distance divided by the flow velocity. The transmission time determines the time delay characteristics of the thermal coupling effect. During transmission, the coolant continuously loses heat to the outside through the pipe walls. The system calculates the proportion of heat attenuation during transmission based on the transmission time and the heat dissipation coefficient per unit length of the pipe. The heat dissipation coefficient per unit length describes the proportion of heat lost per unit length of coolant flow in the pipe relative to the heat it carries; its value is related to the pipe material, pipe wall thickness, pipe insulation, and ambient temperature. For well-insulated liquid-cooled pipes, the heat dissipation coefficient per unit length is relatively small, typically between 0.5% and 2% per meter; for exposed or poorly insulated pipes, the heat dissipation coefficient may be higher. The heat transfer attenuation ratio is calculated using an exponential attenuation model: the attenuation ratio equals 1 minus the negative heat dissipation coefficient per unit length (base of the natural constant) multiplied by the power of the flow distance. The matrix dimension is N x N, where N is the total number of computing nodes in the cluster. The system fills the corresponding matrix positions with the heat transfer attenuation coefficient between any two computing nodes, with the row index representing the heat source node and the column index representing the heat target node. For computing node pairs without a coolant flow path, the system sets their heat transfer attenuation coefficient to zero. Cases where no flow path exists include: the two nodes are located on different cooling branches, and the coolant will not flow from one branch to the other; or they are on the same branch, but the target node is upstream of the source node, and the coolant flow direction is opposite to the heat transfer direction.The thermal coupling strength matrix comprehensively describes the thermal coupling relationship between any two nodes in a liquid-cooled computing cluster.

[0080] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides an adaptive task splitting pipeline optimization system, which includes: a heat conduction parameter set establishment module, a performance trajectory calculation module, and a pipeline execution control module, with each module connected in sequence via communication.

[0081] The heat conduction parameter set establishment module is used to obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connection relationship, computing node location, coolant flow rate and radiator capacity; and establishes the heat conduction parameter set of each computing node based on the physical topology.

[0082] The performance trajectory calculation module is used to collect real-time operating data of computing nodes and cooling systems, extract the current chip temperature and power consumption of each computing node, and the real-time liquid temperature of each key monitoring point of the cooling system from the operating data; determine the time delay constant of each computing node based on the heat conduction parameter set, and set the length of the prediction time window based on the time delay constant; obtain the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue, and calculate the heat power release curve of each task unit; superimpose the heat power release curves to generate the expected heat load input sequence of each computing node within the prediction time window; combine the expected heat load input sequence with the current chip temperature, current inlet liquid temperature and time delay constant of each computing node to calculate the temperature trajectory of each computing node step by step through the heat transfer equation; compare the temperature trajectory with the frequency reduction trigger temperature threshold of each computing node to obtain the degree of temperature exceedance, and infer the available computing frequency at the corresponding time as the performance trajectory of each computing node based on the degree of temperature exceedance.

[0083] The pipeline execution control module is used to generate a new task splitting decision after coordinating intervention on the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction; and to control pipeline execution based on the new task splitting decision.

[0084] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0085] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive task splitting pipeline optimization method, characterized in that, The method includes: Obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connection relationship, computing node location, coolant flow rate and radiator capacity; establish a set of heat conduction parameters for each computing node based on the physical topology; The system collects real-time operational data from computing nodes and the cooling system, extracting the current chip temperature and power consumption of each computing node, as well as the real-time liquid temperature of each key monitoring point in the cooling system. It determines the time delay constant of each computing node based on the heat conduction parameter set and sets the length of the prediction time window based on this time delay constant. It acquires the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue and calculates the heat release curve of each task unit. The heat release curves are superimposed to generate the expected heat load input sequence for each computing node within the prediction time window. The expected heat load input sequence, combined with the current chip temperature, current inlet liquid temperature, and time delay constant of each computing node, is used to iteratively calculate the temperature trajectory of each computing node step-by-step through the heat transfer equation. The temperature trajectory is compared with the frequency reduction trigger temperature threshold of each computing node to obtain the degree of temperature exceedance. Based on the degree of temperature exceedance, the available computing frequency at the corresponding moment is inferred as the performance trajectory of each computing node. When it is determined from the performance trajectory that a computing node will trigger frequency reduction, a new task splitting decision is generated after collaborative intervention on the unexecuted task flow; the pipeline execution is then adjusted according to the new task splitting decision.

2. The adaptive task splitting pipeline optimization method according to claim 1, characterized in that, The method for determining the time delay constant of each computing node based on the set of heat conduction parameters includes: The heat capacity parameters of each computing node and the thermal resistance parameters between the computing node and the coolant are obtained from the heat conduction parameter set; the inherent thermal time constant of each computing node is calculated based on the heat capacity parameters and thermal resistance parameters. Obtain the current inlet liquid temperature and coolant flow rate of each computing node from the operating data, calculate the convective heat transfer coefficient of the coolant based on the coolant flow rate, and determine the equivalent thermal resistance value under the current heat dissipation conditions in combination with the current inlet liquid temperature. The dynamic thermal time constant of each computing node under the current operating conditions is calculated based on the equivalent thermal resistance and heat capacity parameters; the time delay constant of each computing node is obtained by weighted fusion of the dynamic thermal time constant and the inherent thermal time constant.

3. The adaptive task splitting pipeline optimization method according to claim 1, characterized in that, The method for superimposing the heat power release curves to generate the expected heat load input sequence for each computing node within the prediction time window includes: Based on the pipeline task scheduling plan, obtain the start execution time and expected duration of each task unit on each computing node, align the heat power release curve of each task unit to a unified time axis according to the start execution time, and accumulate the heat power release curves of all task units to be executed on the same computing node moment by moment to obtain the local heat load sequence of the computing node. Based on the connection relationship of the liquid cooling loop, determine the set of upstream computing nodes for each computing node, calculate the contribution of the local heat load sequence of each computing node in the upstream computing node set to the temperature rise of the coolant, and then time-shift the temperature rise contribution according to the transmission delay of the coolant flowing from the upstream computing node to the current computing node and add it to the local heat load sequence of the current computing node to obtain the expected heat load input sequence of each computing node within the prediction time window.

4. The adaptive task splitting pipeline optimization method according to claim 1, characterized in that, The method for generating a new task splitting decision after collaboratively intervening in the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction includes: Extract the algorithm type and batch size of each scheduled task unit from the unexecuted task flow. Determine the ratio of computational intensity to memory access intensity of the task unit based on the algorithm type. Determine the parallelism and continuous execution duration of the task unit based on the batch size. Combine the ratio, parallelism, and continuous execution duration with historical execution records to calculate the heat density index and heat persistence index of each task unit. Based on the heat density index and heat persistence index, classify the task units in the unexecuted task flow into high heat impact tasks and low heat impact tasks. Based on the connection relationship between high thermal shock tasks and liquid cooling circuits, a spatiotemporal orchestration result is obtained through distributed orchestration. Based on the margin space between the predicted available computing frequency and the frequency reduction trigger temperature threshold in the performance trajectory of each computing node, the execution frequency of high thermal shock tasks is actively constrained and optimized to determine the execution frequency configuration that balances the task completion time and the thermal power release rate. The spatiotemporal orchestration result and the execution frequency configuration are integrated to generate a new task splitting decision.

5. The adaptive task splitting pipeline optimization method according to claim 4, characterized in that, The method for classifying task units in an unexecuted task flow into high-thermal-shock tasks and low-thermal-shock tasks based on heat density and thermal sustainability indices includes: The heat density index and heat sustainability index of each task unit are used to construct a two-dimensional feature vector. The heat density index and heat sustainability index of historical tasks in the liquid-cooled computing cluster are collected as a clustering sample set. The clustering sample set is used as a similarity measure by the weighted Euclidean distance between the heat density index and the heat sustainability index. The clustering centers are iteratively updated until convergence, resulting in the first clustering center representing high heat load characteristics and the second clustering center representing low heat load characteristics. Calculate the distance between the two-dimensional feature vector of each task unit to be scheduled and the first and second cluster centers, respectively. Compare the distance of the same task unit to the first cluster center with the distance to the second cluster center. Task units whose distance to the first cluster center is less than their distance to the second cluster center are classified as high thermal shock tasks, and task units whose distance to the second cluster center is less than their distance to the first cluster center are classified as low thermal shock tasks. Calculate the thermal load margin of each cooling branch based on the heat dissipation capacity of each cooling branch in the liquid cooling circuit connection relationship and the current liquid temperature state. For task units in the cluster boundary region in the classification results, perform a secondary judgment based on the thermal load margin of the cooling branch to which their target calculation node belongs. Boundary task units with insufficient thermal load margin are reclassified as high thermal shock tasks.

6. The adaptive task splitting pipeline optimization method according to claim 5, characterized in that, The method of using a weighted Euclidean distance between heat density and heat persistence indices as a similarity measure for clustered sample sets includes: The number of series nodes and the coolant flow rate of each cooling branch are obtained based on the connection relationship of the liquid cooling circuit. Cooling branches whose ratio of the number of series nodes to the coolant flow rate is greater than a preset sensitivity threshold are marked as thermally coupled sensitive branches. Thermal coupling sensitivity is obtained by calculating the proportion of computing nodes carried by thermally coupled sensitive branches to the total number of computing nodes in the liquid-cooled computing cluster. When the thermal coupling sensitivity is higher than the preset benchmark value, the weight coefficient of the thermal density index is increased and the weight coefficient of the thermal sustainability index is decreased. When the thermal coupling sensitivity is lower than the preset benchmark value, the weight coefficient of the thermal sustainability index is increased and the weight coefficient of the thermal density index is decreased. The weighted Euclidean distance of each two-dimensional feature vector in the clustered sample set is calculated using the adjusted weight coefficients as a similarity measure.

7. The adaptive task splitting pipeline optimization method according to claim 4, characterized in that, The method for obtaining spatiotemporal orchestration results by dispersing and orchestrating based on the connection relationship between high thermal shock tasks and liquid cooling circuits includes: Based on the connection relationship of the liquid cooling circuit, the topological position of each computing node on the coolant flow path is extracted, the heat transfer attenuation coefficient between any two computing nodes is calculated, and the thermal coupling strength matrix is ​​constructed. High thermal shock tasks and available computing nodes are regarded as two types of vertices in a bipartite graph, and the reciprocal of the elements in the thermal coupling strength matrix is ​​used as the edge weight to construct a task node allocation bipartite graph. The task node allocation bipartite graph is used to solve the spatial allocation scheme by performing the maximum weight matching algorithm; the coolant delivery delay is determined according to the upstream and downstream positional relationship of the computing nodes allocated to each high thermal shock task; the execution time of tasks on the same cooling branch is staggered and arranged with the coolant delivery delay as a constraint; the spatial allocation scheme and the staggered arrangement together constitute the spatiotemporal arrangement result.

8. The adaptive task splitting pipeline optimization method according to claim 7, characterized in that, The method for calculating the heat transfer attenuation coefficient between any two computing nodes and constructing the thermal coupling strength matrix includes: The coolant flow path between any two computing nodes is traversed according to the liquid cooling loop connection relationship. The physical length of each pipe segment on the path is accumulated to obtain the coolant flow distance. The transmission time of the coolant from the upstream computing node to the downstream computing node is calculated based on the coolant flow distance and the current coolant flow rate. The heat attenuation ratio of the coolant during the transmission process is calculated based on the transmission time and the heat dissipation coefficient per unit length of the pipe as the heat transfer attenuation coefficient. A matrix framework is constructed using each computing node in the liquid-cooled computing cluster as the row index and column index. The heat transfer attenuation coefficient between any two computing nodes is filled into the corresponding matrix position. For computing node pairs without a coolant flow path, their heat transfer attenuation coefficient is set to zero to obtain the thermal coupling strength matrix.

9. An adaptive task splitting pipeline optimization system, characterized in that, The system includes: a heat conduction parameter set establishment module, a performance trajectory calculation module, and a pipeline execution control module, with each module connected in a sequential communication manner. A heat conduction parameter set establishment module is used to obtain the physical topology of the liquid-cooled computing cluster, which includes the liquid cooling loop connection relationship, computing node location, coolant flow rate and radiator capacity; and to establish the heat conduction parameter set of each computing node based on the physical topology. The performance trajectory calculation module is used to collect real-time operating data of computing nodes and cooling systems, extract the current chip temperature and power consumption of each computing node, and the real-time liquid temperature of each key monitoring point of the cooling system from the operating data; determine the time delay constant of each computing node based on the heat conduction parameter set, and set the length of the prediction time window based on the time delay constant; obtain the estimated computational load and historical power consumption characteristics of each task unit in the current in-transit task queue, and calculate the heat power release curve of each task unit; superimpose the heat power release curves to generate the expected heat load input sequence of each computing node within the prediction time window; combine the expected heat load input sequence with the current chip temperature, current inlet liquid temperature and time delay constant of each computing node to iteratively calculate the temperature trajectory of each computing node step by step through the heat transfer equation; compare the temperature trajectory with the frequency reduction trigger temperature threshold of each computing node to obtain the degree of temperature exceedance, and infer the available computing frequency at the corresponding time as the performance trajectory of each computing node based on the degree of temperature exceedance. The pipeline execution control module is used to generate a new task splitting decision after coordinating intervention on the unexecuted task flow when it is determined from the performance trajectory that a computing node will trigger frequency reduction; and to control pipeline execution based on the new task splitting decision.

Images