Load migration method and device

Abstract

The embodiment of the invention provides a load migration method and device, and is applied to the technical field of cross-cloud computing power pools. The method comprises the following steps: at a target moment, determining a source node needing load migration and candidate nodes with possible load migration from a plurality of nodes of a cross-cloud computing power pool; and determining whether migration is triggered in a future preset time period or not according to the popularity value sequence of the source node at the target moment, the popularity value change sequence, the popularity value of the candidate node at the target moment and the link bandwidth from the source node to the candidate node. And if the migration task is triggered, determining a migrated target node from the candidate nodes according to the heat value, the instant electricity price and the link bandwidth of the candidate nodes at the target moment, and migrating the workload to the target node. In this way, the determination accuracy of the trigger migration can be improved, and then the resource utilization rate and the operation and maintenance reliability of the cross-cloud computing power pool are improved.

Description

Technical Field

[0001] This application relates to the field of cross-cloud computing power pool technology, and in particular to a load migration method and apparatus. Background Technology

[0002] A cross-cloud computing power pool is a computing infrastructure that integrates public clouds, private clouds, and edge nodes. It is used to support artificial intelligence inference and large-scale offline tasks. To achieve elastic scheduling of cross-cloud resources and ensure the efficient and stable operation of various tasks, workloads need to be migrated from resource-constrained nodes to resource-sufficient nodes without interrupting business operations. This balances the overall resource utilization of the cross-cloud computing power pool and avoids task delays or service interruptions caused by single-point overload.

[0003] Currently, migration is often triggered based on fixed thresholds. Specifically, it monitors metrics such as GPU utilization, CPU utilization, bandwidth utilization, and latency of nodes, and directly triggers migration when one or more metrics reach an upper limit threshold. For example, migration is directly triggered when the GPU utilization of a node is greater than or equal to 80%.

[0004] However, this fixed threshold-triggered migration method has the following problems: the resource status, link quality, and service load of nodes in the cross-cloud computing power pool all exhibit minute-level tidal fluctuations. Fixed thresholds cannot respond to these dynamic changes in real time, resulting in insufficient accuracy in determining the trigger for migration. For example, during peak hours, a node's short-term metric reaching the upper limit threshold may be a momentary spike, in which case migration is unnecessary. During off-peak hours, a node's metric approaching the upper limit threshold may already indicate resource overload, requiring migration to be triggered. Insufficient accuracy in determining the trigger for migration will affect the resource utilization and operational reliability of the cross-cloud computing power pool. Summary of the Invention

[0005] This application provides a load migration method and apparatus to improve the accuracy of triggering migration, thereby improving the resource utilization and operational reliability of cross-cloud computing power pools.

[0006] In a first aspect, embodiments of this application provide a load migration method applied to a cross-cloud computing power pool, wherein the cross-cloud computing power pool includes multiple nodes, and the method includes:

[0007] At the target time, the source node and candidate node of the plurality of nodes are determined; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated.

[0008] Based on the heat value sequence and heat value change sequence of the source node at the target time, as well as the heat value and link bandwidth of the candidate node at the target time, determine whether to trigger a load migration task in a future preset period of time;

[0009] Wherein, the link bandwidth of the candidate node represents the link bandwidth from the source node to the candidate node, the future preset time period is the time period after the target time; the heat value sequence is a sequence composed of the heat values ​​of the source node within the target time period, the target time period is the time period before the target time; the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence;

[0010] If a load migration task is triggered, a load migration strategy is determined based on the popularity value, real-time electricity price, and link bandwidth of the candidate node at the target time, and load migration is performed according to the load migration strategy; the load migration strategy includes a target node, which is one of the candidate nodes.

[0011] Optionally, determining whether to trigger a load migration task in a future preset time period based on the heat value sequence and heat value change sequence of the source node at the target time, and the heat value and link bandwidth of the candidate node at the target time, includes:

[0012] Obtain the first encoding vector corresponding to the heat value sequence, and the second encoding vector corresponding to the heat value change sequence;

[0013] The first encoding vector, the second encoding vector, the heat value of the candidate node at the target time, and the link bandwidth are concatenated to obtain the state vector.

[0014] The state vector is processed using a migration triggering model to determine whether a load migration task will be triggered within a preset future time period. The input of the migration triggering model is the state vector, and the output of the migration triggering model is a Boolean flag. When the Boolean flag is a first flag, the load migration task is determined to be triggered. When the Boolean flag is a second flag, the load migration task is paused.

[0015] Optionally, the transfer triggering model includes a decision network and a value network; the transfer triggering model is a trained model, and during the training process, the transfer triggering model adjusts the decision parameters of the decision network according to the state value obtained by the value network;

[0016] The step of processing the state vector using a migration triggering model to determine whether a load migration task will be triggered within the preset future time period includes:

[0017] The state vector is processed by the trained decision network to obtain the transition probability;

[0018] If the migration probability is greater than or equal to the probability threshold, output the first flag; if the migration probability is less than the probability threshold, output the second flag.

[0019] Optionally, the migration-triggered model is trained using a proximal policy optimization algorithm, and an immediate reward function is used during the training process;

[0020] The instant reward function is negatively correlated with the reward value of migration costs and positively correlated with the improvement in the achievement rate of Service Level Agreements (SLAs) within the preset future time period.

[0021] Optionally, determining the source node and candidate nodes of the plurality of nodes includes:

[0022] Obtain the popularity value of each of the plurality of nodes at the target time; the popularity value of a node includes the popularity value corresponding to each service in the node;

[0023] Based on the heat value of each node at the target time, source nodes and candidate nodes are determined from the plurality of nodes; the highest heat value of the source node at the target time is greater than or equal to an overheating threshold, the lowest heat value of the candidate node at the target time is less than or equal to a slightly cold threshold, and the overheating threshold is greater than the slightly cold threshold; the highest heat value of the source node is the maximum value of the heat value corresponding to each service in the source node, and the lowest heat value of the candidate node is the minimum value of the heat value corresponding to each service in the candidate node.

[0024] Optionally, determining the load migration strategy based on the candidate node's popularity value, real-time electricity price, and link bandwidth at the target time includes:

[0025] Using a graph attention network, the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time are processed to determine the score of the candidate nodes; the score of the candidate nodes is used to describe the relative priority of migrating load to the candidate nodes;

[0026] The load migration strategy is determined based on the scores of the candidate nodes.

[0027] Optionally, the load migration strategy further includes a target flow-switching time slot, the target flow-switching time slot indicating the time slot for load migration, and the method further includes:

[0028] The preset future time period is cut into streams according to preset rules to obtain N stream cutting time slots; where N is an integer greater than or equal to 1.

[0029] The graph attention network is used to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes, including:

[0030] Using the graph attention network, the heat value, instantaneous electricity price, link bandwidth, and N flow switching time slots of the candidate nodes at the target time are processed to determine the scores of the N candidate nodes; the scores of the candidate nodes correspond one-to-one with the flow switching time slots.

[0031] Optionally, before using a graph attention network to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate node at the target time to determine the score of the candidate node, the method includes:

[0032] Obtain a candidate graph set; the candidate graph set includes vertices, vertex attributes, edges, and edge attributes; the vertices include a first vertex corresponding to the source node and a second vertex corresponding to the candidate node; the vertex attributes include the heat value, instantaneous electricity price, and available processor capacity of the candidate node corresponding to the second vertex at the target time; the edges include a one-way edge from the first vertex to the second vertex; the edge attributes include the link bandwidth and round-trip latency from the source node to the candidate node.

[0033] The step of using a graph attention network to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes includes: processing the candidate graph set using a graph attention network to determine the score of the candidate nodes.

[0034] Optionally, after determining the load migration strategy, the method further includes:

[0035] Determine the real-time resources and real-time link bandwidth of the target node;

[0036] If the real-time resources are greater than or equal to a preset resource threshold, and the real-time link bandwidth is greater than or equal to a preset bandwidth threshold, load migration is performed according to the load migration strategy.

[0037] Secondly, embodiments of this application provide a load migration device applied to a cross-cloud computing power pool, wherein the cross-cloud computing power pool includes multiple nodes, and the device includes:

[0038] The first determining unit is used to determine the source node and candidate node of the plurality of nodes at a target time; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated.

[0039] The second determining unit is used to determine whether to trigger a load migration task in a future preset period based on the heat value sequence and heat value change sequence of the source node at the target time, as well as the heat value and link bandwidth of the candidate node at the target time.

[0040] Wherein, the link bandwidth of the candidate node represents the link bandwidth from the source node to the candidate node, the future preset time period is the time period after the target time; the heat value sequence is a sequence composed of the heat values ​​of the source node within the target time period, the target time period is the time period before the target time; the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence;

[0041] The third determining unit is used to determine a load migration strategy based on the popularity value, real-time electricity price and link bandwidth of the candidate node at the target time if a load migration task is triggered, so as to perform load migration according to the load migration strategy; the load migration strategy includes a target node, which is one of the candidate nodes.

[0042] Thirdly, embodiments of this application provide an electronic device, including:

[0043] Memory, used to store computer programs;

[0044] A processor for executing the computer program to implement the method as described in any one of the first aspects.

[0045] Fourthly, embodiments of this application provide a computer program that, when run on a computer, causes the computer to perform the method in any of the possible implementations of any of the above aspects.

[0046] Fifthly, embodiments of this application provide a computer-readable storage medium storing a computer program (also referred to as code or instructions) that, when run on a computer, causes the computer to perform the method in any of the possible implementations of any of the above aspects.

[0047] Sixthly, embodiments of this application provide a chip system including one or more processors for calling and executing instructions stored in memory, causing the methods in any of the above aspects or possible implementations to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0048] This application provides a load migration method and apparatus. The method includes: at a target time, determining the source node requiring load migration and candidate nodes that may be migrated from multiple nodes in a cross-cloud computing power pool; and determining whether to trigger migration within a preset time period based on the source node's heat value sequence, heat value change sequence, candidate node's heat value at the target time, and link bandwidth from the source node to the candidate node. The heat value sequence is a sequence of source node heat values ​​within the target time period, and the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence. This allows for the capture of dynamic heat value changes through two indicators: the heat value change trend and the heat value change rate of the source node, thereby improving the accuracy of triggering migration. Furthermore, by comprehensively considering the link bandwidth and heat value of candidate nodes, secondary overload or migration interruption after load migration is avoided, further improving the accuracy of triggering migration. Therefore, this method can improve the resource utilization and operational reliability of the cross-cloud computing power pool.

[0049] Furthermore, if a migration task is triggered, the target node for migration is determined from the candidate nodes based on their popularity, real-time electricity price, and link bandwidth at the target time, and the workload is migrated to that target node. By comprehensively considering the popularity, real-time electricity price, and link bandwidth of candidate nodes at the target time when determining where to migrate, it is possible to migrate the load of the source node to the target node with low cost, high reliability, and high migration rate.

[0050] Furthermore, the load migration method provided in this application decouples whether migration is triggered and where to migrate, that is, it divides whether migration is triggered and where to migrate into two stages and determines them separately. Compared with a single stage that directly determines where to migrate and whether migration is triggered, it can shorten the determination time and improve scheduling efficiency. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram of a cross-cloud computing power pool architecture provided for an embodiment of this application;

[0053] Figure 2 This is a schematic diagram of the architecture of a load migration system provided in an embodiment of this application;

[0054] Figure 3 A flowchart of a load migration method provided in an embodiment of this application;

[0055] Figure 4 This is a schematic diagram of another load migration system provided in an embodiment of this application;

[0056] Figure 5 An interaction diagram of a load migration method provided in an embodiment of this application;

[0057] Figure 6 A flowchart illustrating a method for obtaining the heat risk coefficient of a node based on an exponential memory-logistic curve, as provided in this application embodiment;

[0058] Figure 7 A schematic diagram of the heat risk coefficient curves corresponding to the three service levels provided in the embodiments of this application;

[0059] Figure 8 A schematic diagram illustrating a method for obtaining a target flow-cutting time slot and a target node, provided in an embodiment of this application;

[0060] Figure 9 This is a schematic diagram of a load migration device provided in an embodiment of this application. Detailed Implementation

[0061] To enable those skilled in the art to better understand the present application, the technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0063] This application provides a load migration method that can be applied to cross-cloud computing power pools. Cross-cloud computing power pools integrate multiple nodes such as private clouds, public clouds, and edge nodes to support artificial intelligence inference and large-scale offline tasks.

[0064] For example, participants Figure 1 The diagram shows an architecture of a cross-cloud computing power pool according to an embodiment of this application. The cross-cloud computing power pool 10 includes N1 private clouds, N2 public clouds, and N3 edge nodes. N1, N2, and N3 are all positive integers. For example, Figure 1The demonstration included two private clouds, one public cloud, and two edge nodes. The two private clouds were designated as Private Cloud 1 and Private Cloud 2, and the two edge nodes were designated as Edge Node 1 and Edge Node 2.

[0065] In this embodiment of the application, workload migration can be performed between private clouds, between public clouds, between edge nodes, and between public clouds, private clouds, and edge nodes. For example, Figure 1 The demonstration showed that workloads in private cloud 1 can be migrated to public cloud, edge node 1, and edge node 2; workloads in public cloud can be migrated to private cloud 2 and edge node 1; and workloads in edge node 1 can be migrated to public cloud, private cloud 1, and private cloud 2.

[0066] The load migration method provided in this application first determines a source node and candidate nodes from multiple nodes in a cross-cloud computing pool at a target time. The source node is the node requiring load migration, and the candidate node is the node to which the load may be migrated. Then, based on the source node's popularity value sequence, popularity value change sequence, candidate node's popularity value at the target time, and the link bandwidth from the source node to the candidate node, it is determined whether migration should be triggered within a preset future time period. The popularity value sequence is a sequence of source node popularity values ​​within the target time period, and the popularity value change sequence is a sequence obtained by first-order difference of the popularity value sequence. The popularity value sequence reflects the trend of popularity value changes in the source node, while the popularity value change sequence reflects the rate of change of the source node's popularity value. By comprehensively considering both the popularity value sequence and the popularity value change sequence, the dynamic trend of the source node's popularity value can be accurately captured, thereby accurately determining whether migration should be triggered.

[0067] For example, if the target time is the current time, then for Figure 1The cross-cloud computing pool shown has the following candidate nodes: if the source node is private cloud 1, the candidate nodes are public cloud, edge node 1, and edge node 2. The heat value sequence of private cloud 1 at the target time is [75, 76, 77, 79, 80, 82, 83, 85, 86, 88, 89, 90, 91, 92, 93, 92, 91, 93, 95, 96, 97, 98, 99, 100, 99, 98, 97, 96, 95, 94]. The heat value change sequence of private cloud 1 at the target time is [1, 1, 2, 1, 2, 1, 2, 1, 2, 1, 1, 1, 1, -1, -1, 2, 2, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1, -1, -1, -1]. By using curve fitting, such as the Logistic growth model, the popularity value of the private cloud in a future preset time period, such as the next 5 minutes, can be obtained to be between 98 and 102. Assuming a migration popularity threshold (called the overheat threshold) is set, for example, 98, it can be determined that migration needs to be triggered within the next 5 minutes. Furthermore, by combining the popularity value of the public cloud at the target time (for example, 15), the link bandwidth from private cloud 1 to the public cloud (for example, 500MB / s), the popularity value of edge node 1 at the target time (for example, 18), the link bandwidth from private cloud 1 to edge node 1 (for example, 300MB / s), the popularity value of edge node 2 at the target time (for example, 12), and the link bandwidth from private cloud 1 to edge node 2 (for example, 400MB / s), it can be further determined whether migration should be triggered within the next W=5 minutes. For example, the ability of a node to take over the load of the source node can be set, i.e., the node cold threshold is 20, the bandwidth threshold is 200MB / s, the heat value of the candidate nodes is less than or equal to the cold threshold, and the link bandwidth is greater than or equal to the bandwidth threshold. This means that the candidate nodes meet the requirements of low heat and high bandwidth, providing feasibility support for migration triggering, and thus determining to trigger migration.

[0068] This approach can capture the dynamic trend of heat value through two indicators: the heat value change trend and the heat value change rate of the source node, thereby improving the accuracy of determining the trigger migration. By comprehensively considering the link bandwidth and heat value of the candidate node, it can avoid secondary overload or migration interruption after load migration, further improving the accuracy of determining the trigger migration. Thus, this method can improve the resource utilization and operation and maintenance reliability of cross-cloud computing power pools.

[0069] Furthermore, if a migration task is triggered, the target node for migration is determined from the candidate nodes based on the candidate nodes' popularity value, real-time electricity price, and link bandwidth at the target time, and the workload is migrated to the target node. For example, continuing with the example above, if the real-time electricity price for the public cloud at the target time is 1 yuan / kWh, the real-time electricity price for edge node 1 at the target time is 0.5 yuan / kWh, and the real-time electricity price for edge node 2 at the target time is 0.8 yuan / kWh, triggering a migration task allows the selection of a target node from the candidate nodes. This target node can simultaneously achieve high load capacity, low migration cost, and high migration efficiency. For example, the target node might be edge node 2, which has a lower popularity value, higher link bandwidth, and a relatively lower real-time electricity price compared to public cloud services. In other words, edge node 2 can simultaneously achieve high load capacity, low migration cost, and high migration efficiency.

[0070] When determining the destination, by comprehensively considering the popularity of candidate nodes at the target time, real-time electricity price, and link bandwidth, it is possible to migrate the load from the source node to the target node with low cost, high reliability, and high efficiency.

[0071] Furthermore, in this embodiment of the application, the method decouples the migration timing and the migration location, dividing them into two stages that are executed separately. Compared to directly determining the migration location and timing in a single stage, this can shorten the determination time and improve scheduling efficiency.

[0072] The load migration method provided in this application embodiment can be executed by a load migration system (or load migration platform). The load migration system can perform the following operations: at a target time, determine the source node requiring load migration and the candidate nodes to which the load may be migrated from multiple nodes in the cross-cloud computing power pool; based on the source node's heat value sequence, heat value change sequence, candidate node's heat value at the target time, and link bandwidth from the source node to the candidate node, determine whether to trigger migration within a preset future time period. This allows for precise capture of the dynamic trend of heat value changes through two indicators: the heat value change trend and the heat value change rate of the source node. Compared to triggering migration using a fixed threshold, this improves the accuracy of determining migration triggers. Furthermore, by comprehensively considering the link bandwidth and heat value of candidate nodes, it avoids secondary overload or migration interruption after load migration, further improving the accuracy of determining migration triggers. By improving the accuracy of determining migration triggers, the resource utilization and operational reliability of the cross-cloud computing power pool are improved.

[0073] Furthermore, if the load migration system triggers a migration task, it can determine the target node from among the candidate nodes based on the candidate nodes' popularity, real-time electricity price, and link bandwidth at the target time, and then migrate the workload to that target node. By comprehensively considering the candidate nodes' popularity, real-time electricity price, and link bandwidth at the target time when determining the migration destination, the system can achieve low-cost, high-reliability, and high-speed migration of workloads from the source node to the target node. Moreover, by decoupling the migration time from the migration destination to the migration location, this method can shorten the determination time compared to a single-stage direct determination of the migration location and timing, thereby improving scheduling efficiency.

[0074] In practical applications, a load migration system can include a software system, which can be provided to the user as a software package for self-deployment, such as on a local physical server or in a private cloud. In some possible implementations, the load migration system can also be deployed in a public cloud and provided to users as a cloud service. For example, cloud service providers can offer users a one-stop system service integrating the functions of the aforementioned core components.

[0075] To facilitate understanding, the following is in conjunction with the appendix. Figure 2 The system architecture of the load migration system provided in the embodiments of this application is described.

[0076] join Figure 2 The diagram shown is an architectural schematic of a load migration system provided in an embodiment of this application. The load migration system 20 includes a heat analyzer 201, a scheduler 202, and an orchestrator 203. The heat analyzer 201 is communicatively connected to the scheduler 202, and the scheduler 202 is communicatively connected to the orchestrator 203.

[0077] The heat analyzer 201 is used to determine the heat value of the source node and the heat value of the candidate node at the target time, and sends the determined heat value to the scheduler 202.

[0078] Scheduler 202 is used to determine, at a target time, the source node and candidate nodes of multiple nodes across the cloud computing pool, and, based on the source node's popularity value sequence, popularity value change sequence, candidate node's popularity value at the target time, and link bandwidth from the source node to the candidate node, determine whether to trigger migration within a preset time period in the future. Scheduler 202 is also used to determine a load migration strategy based on the candidate node's popularity value at the target time, real-time electricity price, and link bandwidth when migration is determined to be triggered. The load migration strategy includes a target node, which is one of the candidate nodes. Scheduler 202 is used to send the load migration strategy to orchestrator 203.

[0079] The orchestrator 203 is used to migrate the workload of the source node to the target node according to the load migration strategy.

[0080] The following is in conjunction with the appendix Figures 3-8 The load migration method provided in the embodiments of this application will be described. This load migration method can be applied to... Figure 2 The load migration system shown is used as an example for detailed explanation.

[0081] Appendix Figure 3 A flowchart illustrating a load migration method provided in this application embodiment. The method specifically includes the following steps:

[0082] S10, at the target time, determine the source node and candidate node among multiple nodes in the cross-cloud computing power pool; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated.

[0083] There can be one or more source nodes, which is not specifically limited in this embodiment. There can also be one or more candidate nodes, which is not specifically limited in this embodiment.

[0084] In some examples, it's possible to manually specify which nodes among multiple nodes are source nodes and which are candidate nodes. For example, for... Figure 1 As shown, operations and maintenance personnel can manually configure the source node and candidate node in the cross-cloud computing power pool through the load migration system. For example, the source node is private cloud 1, and the candidate nodes are public cloud and edge node 2.

[0085] In other examples, the load migration system can obtain the heat value of each node at a target time. The node's heat value describes its "hot" or "cold" status, reflecting its load pressure and resource scarcity. A higher heat value indicates greater load pressure and resource scarcity, necessitating the migration of that node's workload to other nodes for load balancing. The load migration system can determine source and candidate nodes based on the heat value of each node at the target time; for example, if the source node's heat value at the target time is greater than or equal to a preset heat threshold, and the candidate node's heat value at the target time is less than or equal to a preset heat threshold. For example, for... Figure 1 As shown, if the popularity value of private cloud 1 at the target time is 92, the popularity value of private cloud 2 at the target time is 65, the popularity value of public cloud at the target time is 28, the popularity value of edge node 1 at the target time is 22, and the popularity value of edge node 2 at the target time is 35, and the preset popularity threshold is 80, then the source node is private cloud 1, and the other nodes are candidate nodes of private cloud 1. This method does not require manual intervention; it only judges based on the popularity value of the nodes.

[0086] In some examples, a node's popularity value is related to the importance level of a business. For instance, a node may include M businesses, where M is a positive integer (e.g., M=3). The node's popularity value includes the popularity values ​​of all M businesses, and these popularity values ​​are related to the importance level of the business. Specifically, the popularity value corresponding to the first business level is lower than that corresponding to the second business level, and the popularity value corresponding to the second business level is lower than that corresponding to the third business level. The first business level is considered more critical than the second business level, and the second business level is considered more critical than the third business level. The load migration system obtains the lowest and highest popularity values ​​from the node's popularity values. If the lowest popularity value of a node is lower than a "cold" threshold, the node is identified as a candidate node. If the highest popularity value of a node is greater than or equal to an "overheated" threshold, the node is identified as the source node. The "overheated" threshold is greater than the "cold" threshold. This approach narrows down the range of candidate nodes and accurately adapts to differences in business levels, avoiding the mis-migration of critical businesses. For example, if private cloud 1 includes three services, namely service 1, service 2, and service 3, where service 1 is the first service level, service 2 is the second service level, and service 3 is the third service level, the load migration system can obtain the heat value of service 1 at the target time, for example, 65, the heat value of service 2 at the target time, for example, 78, and the heat value of service 3 at the target time, for example, 93. Edge node 2 includes service 4 and service 5, where the heat value of service 5 at the target time is 45, and the heat value of service 4 at the target time is 27. By setting an overheat threshold of 90 and a cold threshold of 30, it can be determined that the highest heat value of private cloud 1 is greater than the overheat threshold, making private cloud 1 the source node; the lowest heat value of edge node 2 is lower than the cold threshold, making edge node 1 a candidate node. Part or all of service 3 in private cloud 1 can be migrated to edge node 1.

[0087] It should be noted that, in the embodiments of this application, the overheat threshold and the underheat threshold can be adjusted as needed. For example, the overheat threshold and the underheat threshold can be updated according to time period, electricity price and / or date, etc., which is not limited in the embodiments of this application.

[0088] In the embodiments of this application, the popularity value of each business level ID in each node can be obtained directly based on the real-time resource consumption index of the business and according to the weight allocated to the business level. Alternatively, it can be achieved in other ways, which are not limited in the embodiments of this application.

[0089] S20: Based on the heat value sequence and heat value change sequence of the source node at the target time, as well as the heat value and link bandwidth of the candidate node at the target time, determine whether to trigger the load migration task in the future preset time period.

[0090] The future preset time period is the time period after the target time. For example, the future preset time period is 5 minutes from now. The target time period is the time period before the target time, for example, 60 seconds before the target time.

[0091] The link bandwidth of a candidate node represents the link bandwidth from the source node to the candidate node, and is used to measure the transmission rate of the available network from the source node to the candidate node.

[0092] The heat value sequence is a sequence composed of the heat values ​​of source nodes within the target time period. In this embodiment, the load migration system acquires the heat value of a node every preset time interval, such as 2 seconds. The load migration system arranges the heat values ​​of the source nodes acquired during the target time period in chronological order to obtain the heat value sequence. The target time period is 60 seconds before the target time. The heat value of the source node is obtained every 2 seconds. The heat values ​​of the source node obtained in chronological order are 75, 76, 77, 79, 80, 82, 83, 85, 86, 88, 89, 90, 91, 92, 93, 92, 91, 93, 95, 96, 97, 98, 99, 100, 99, 98, 97, 96, 95, 94. The heat value sequence of the source node is [75, 76, 77, 79, 80, 82, 83, 85, 86, 88, 89, 90, 91, 92, 93, 92, 91, 93, 95, 96, 97, 98, 99, 100, 99, 98, 97, 96, 95, 94].

[0093] The heat value change sequence is obtained by first-order difference of the heat value sequence. For example, if the heat value sequence of the source node is [75, 76, 77, 79, 80, 82, 83, 85, 86, 88, 89, 90, 91, 92, 93, 92, 91, 93, 95, 96, 97, 98, 99, 100, 99, 98, 97, 96, 95, 94], the heat value change sequence obtained after first-order difference is [1, 1, 2, 1, 2, 1, 2, 1, 2, 1, 1, 1, 1, -1, -1, 2, 2, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1, -1, -1]. The heat value change sequence reflects the rate of change of the heat value.

[0094] In this embodiment of the application, the load migration system can determine whether to trigger a load migration task in a future preset period based on the source node's heat value sequence and heat value change sequence at the target time, as well as the candidate node's heat value and link bandwidth at the target time.

[0095] In some examples, the load migration system can encode the heat value sequence and the heat value change sequence separately, resulting in a first encoding vector and a second encoding vector. For instance, the load migration system uses a gated recurrent unit (GRU) to transform the sequence. The GRU can capture long-term dependencies and change patterns in time-series data and reduce computational complexity.

[0096] The load migration system then concatenates the first encoding vector, the second encoding vector, the heat value of the candidate node at the target time, and the link bandwidth to obtain the state vector. In this embodiment, considering that there may be multiple candidate nodes, the heat value of the candidate node at the target time can be specifically limited to the average of the lowest heat values ​​of multiple candidate nodes, and the link bandwidth of the candidate node at the target time can be specifically limited to the median of the link bandwidths of multiple candidate nodes.

[0097] Furthermore, in this embodiment, the state vector may also include the number of candidate nodes.

[0098] The load migration system can use a migration triggering model to process the state vector and determine whether to trigger a migration task within a preset time period. The migration triggering model takes the state vector as input and outputs a Boolean flag. When the Boolean flag is the first flag, a migration task is triggered; otherwise, triggering the migration task is paused. In the specific implementation, the Boolean flag... .in, Migrate (i.e., the first flag) characterizes the triggering migration. Defer (i.e., the second flag) indicates that the trigger migration is paused.

[0099] In this embodiment of the application, the migration triggering model is a trained model, and the training samples of the model include historical state vectors and Boolean flags corresponding to the historical state vectors.

[0100] In one example, the transfer-triggered model includes a decision network (Actor network) and a value network (Critic network). The trained decision network processes the state vector to obtain transfer probabilities and outputs a first flag when the transfer probability is greater than or equal to a probability threshold, and a second flag when the transfer probability is less than the probability threshold. The value network, during training, acquires state values ​​and uses these values ​​to adjust the decision parameters of the decision network, thereby improving the accuracy of the trained decision network.

[0101] In another example, both the decision network and the pre-processor network are lightweight network frameworks, each consisting of a two-layer structure: a fully connected layer and an output layer. In one specific implementation, the output layer uses the ReLU activation function.

[0102] In this embodiment of the application, the migration triggering model can be trained using a proximal policy optimization algorithm to improve training accuracy.

[0103] In some examples, the migration-triggered model uses an immediate reward function. This immediate reward function is negatively correlated with the reward value of migration costs and positively correlated with the improvement in Service Level Agreement (SLA) over a predetermined future period, as shown in formula (1):

[0104] (1)

[0105] in, This is the reward value of the immediate reward function. −MigrateCost(t) is the reward value for migration cost; the higher the migration cost, the lower this reward value. The reward function sets the SLA improvement target for a future time period W. The greater the improvement, the higher the reward. This immediate reward function allows the trained transfer learning model to maximize the SLA achievement rate while minimizing transfer costs.

[0106] It should be noted that in the embodiments of this application, the probability threshold is a threshold set by those skilled in the art as needed, for example, the probability threshold is 0.5.

[0107] S30: If a load migration task is triggered, determine the load migration strategy based on the candidate node's popularity value at the target time, real-time electricity price, and link bandwidth.

[0108] The load migration strategy includes a target node, which is one of the candidate nodes.

[0109] In some examples, graph attention networks can be used to process the popularity value, instantaneous electricity price, and link bandwidth of candidate nodes at a target time to determine the candidate node's score. Based on the candidate node's score, a load migration strategy can be determined. The candidate node's score describes the relative priority of load migration to that candidate node. In practical applications, the higher (or lower) the candidate node's score, the higher the relative priority for load migration to that candidate node. Higher priority means the load is more likely to migrate to that candidate node. For clarity, the following explanation uses the example of a higher candidate node score indicating higher priority for load migration to that candidate node.

[0110] For example, targeting Figure 1 As shown, the candidate nodes are public cloud A, edge node 1, and edge node 2. The heat value, real-time electricity price, and link bandwidth of the three candidate nodes at the target time are shown in Table 1.

[0111] Table 1 shows the popularity, instantaneous electricity price, and link bandwidth of candidate nodes at the target time.

[0112]

[0113] Table 1 shows the popularity values ​​of the three candidate nodes at the target time. The popularity value of edge node 1 > that of the public cloud > the popularity value of edge node 2, indicating that edge node 2 has a higher load-bearing capacity. The real-time electricity prices of the three candidate nodes at the target time are: public cloud > edge node 2 > edge node 1, indicating that the migration cost is lowest for edge node 1 and highest for the public cloud. The link bandwidth of the three candidate nodes at the target time is: public cloud > edge node 2 > edge node 1, indicating that the migration rate is lowest for edge node 1 and highest for the public cloud.

[0114] Furthermore, to improve the processing accuracy of graph attention networks, the data of the three indicators in Table 1 can be normalized and mapped to dimensionless numbers in the interval [0,1].

[0115] Using a graph attention network, the public cloud scored 0.72, edge node 1 scored 0.89, and edge node 2 scored 0.78. Based on these scores, the target node can be identified as edge node 1.

[0116] In this embodiment of the application, by comprehensively considering the real-time electricity price, popularity value and link bandwidth of the candidate nodes, the target node determined can simultaneously take into account low cost, high capacity and high migration efficiency.

[0117] Furthermore, the load migration strategy also includes a target flow-switching time slot. That is, a specific time slot within a preset future period to migrate the load from the source node to the target node.

[0118] In this embodiment, the load migration system can perform flow switching on a preset future time period according to preset rules, resulting in N flow switching time slots. Here, N is greater than or equal to 1 and is an integer, for example, N=4. The preset rules can be time period equal division rules, such as dividing the future window W into 4 flow switching time slots. The load migration system utilizes a graph attention network to simultaneously process the candidate node's popularity value, instantaneous electricity price, and link bandwidth at the target time, as well as the N flow switching time slots, to determine N scores for the candidate nodes, where each candidate node's score corresponds one-to-one with a flow switching time slot. For example, if the preset future time period includes 4 flow switching time slots, the scores for the candidate nodes in the 4 flow switching time slots are obtained. Based on these scores, the target flow switching time slot and the target node for migration are further determined.

[0119] It should be noted that the migration load strategy can also be determined in other ways in the embodiments of this application, and this application is not limited to these methods.

[0120] S40 performs load migration according to the load migration strategy.

[0121] In the specific implementation, after the scheduler 202 obtains the load migration policy, it sends the load migration policy to the orchestrator 203, and the orchestrator 203 performs the load migration according to the following steps:

[0122] Step ①: State Consistency Snapshot Capture. Specifically, at the moment the migration starts, a momentary pause signal is sent to the source node containing the workload to be migrated, freezing its complete running state, memory data, context information, and other state data. A snapshot of this state data is then captured to obtain snapshot data.

[0123] Step 2: Incremental compression.

[0124] Step 3: Adaptive concurrent transmission. Specifically, based on the real-time link bandwidth, latency jitter, and network stability from the source node to the target node, the transmission concurrency, transmission rate, and retransmission strategy are dynamically adjusted to ensure that data is transmitted to the target cloud quickly and reliably.

[0125] Step 4: Target Cloud Preheating. Specifically, during or after data transmission, a runtime environment identical to that of the source node is pre-built on the target node, and the Checkpoint snapshot state is restored to ensure that the business can be started and run immediately after migration.

[0126] Step 5: Restore Traffic Switching. Specifically, after the target node has warmed up, external business traffic is smoothly switched from the source node to the target node. Once the business is running stably, the relevant data and resources on the source node are cleaned up, completing the entire migration process.

[0127] Furthermore, if any step times out or fails verification, an automatic rollback strategy is triggered, and a negative reward is recorded for the current load migration strategy. Scheduler 202 retrains the migration triggering model based on the migration results, actual time consumption, bandwidth, electricity price, and reward value. Through a closed-loop operation of acquiring the load migration strategy, executing the load migration strategy, and providing feedback on the execution results, the load migration strategy can adaptively evolve in an environment where business and network conditions are constantly changing.

[0128] Furthermore, in this embodiment, after the orchestrator 203 obtains the load migration strategy, it can determine the real-time resources and real-time link bandwidth of the target node. If the real-time resources are greater than or equal to a preset resource threshold and the real-time link bandwidth is greater than or equal to a preset bandwidth threshold, load migration is performed according to the load migration strategy. The preset resource threshold and preset bandwidth threshold are pre-set values ​​that can be adjusted as needed. For example, the preset bandwidth threshold is... The preset resource threshold is 4 GPU resources.

[0129] The load migration method provided in this application captures the dynamic trend of heat value changes by using two indicators: the heat value change trend and the heat value change rate of the source node. Compared with using a fixed threshold to trigger migration, this method improves the accuracy of triggering migration. Furthermore, by comprehensively considering the link bandwidth and heat value of candidate nodes, it avoids secondary overload or migration interruption after load migration, further improving the accuracy of triggering migration. Improving the accuracy of triggering migration can enhance the resource utilization and operational reliability of cross-cloud computing pools. Moreover, if the load migration system triggers a migration task, it can determine the target node from the candidate nodes based on their heat value, real-time electricity price, and link bandwidth at the target time, and migrate the workload to that target node. By comprehensively considering the heat value, real-time electricity price, and link bandwidth of candidate nodes at the target time when determining where to migrate, the method can achieve low-cost and high-reliability migration of workloads to the target node. Furthermore, by decoupling whether migration is triggered and where to migrate, this method shortens the determination time and improves scheduling efficiency compared to directly determining where to migrate and whether migration is triggered in a single stage.

[0130] To enable those skilled in the art to better understand the load migration method provided in the embodiments of this application, the following description is provided with specific examples. In the embodiments of this application, the heat value of each node in the cross-cloud computing power pool at the target time is first determined based on multi-dimensional indicators. The target time is time t. In the embodiments of this application, the heat value of each node includes the heat risk coefficient of the business level P in that node at time t. Then, based on the heat risk coefficient of each node, the source node and candidate nodes are determined. This embodiment decouples whether to migrate and where to migrate, dividing it into stage A and stage B. Stage A is used to execute step S20, and stage B is used to execute step S30. The graph attention network uses a two-layer graph attention network, which will be discussed in detail below with reference to the accompanying drawings.

[0131] Appendix Figure 4 This is a schematic diagram of another load migration system provided in an embodiment of this application. The load migration system 20 includes a monitoring agent module 401, an indicator bus module 402, a heat analyzer 201, a scheduler 202, an orchestrator 203, and an experience database 403.

[0132] The monitoring agent module 401 is used to acquire multi-dimensional metrics of each node in the cross-cloud computing power pool and tag the multi-dimensional metrics of each node before sending them to the metric bus module 402. This ensures that the multi-dimensional metrics of each node meet the required standards, guaranteeing that the metric data from multiple nodes across the computing power pool are accurately clustered according to node and business dimensions when aggregated.

[0133] The indicator bus module 402 is used to receive multi-dimensional indicators from each node. To ensure timing integrity, the indicator bus module 402 is also used to rearrange the multi-dimensional indicators of each node according to the received timing sequence. The indicator bus module 402 then sends the rearranged multi-dimensional indicators of each node to the heat analyzer 201.

[0134] The heat analysis unit 201 is used to obtain the multi-dimensional indicator sequence of each node within the scrolling window, and to determine the heat risk coefficient of each node based on the multi-dimensional indicator sequence of each node. The heat analysis unit 201 is also used to send the heat risk coefficient of each node to the scheduler 202.

[0135] Scheduler 202 receives the heat risk coefficient of each node and, based on the heat risk coefficient, determines whether to trigger migration, when to trigger migration, and where to migrate, thus obtaining a load migration strategy. Caller 202 sends the load migration strategy to orchestrator 203.

[0136] The orchestrator 203 is used to perform load migration based on the load migration strategy and store the load migration information in the experience database 403.

[0137] against Figure 4 The load migration system shown in this application also provides a load migration method in this embodiment. Figure 5 The diagram shown is an interaction diagram of a load migration method provided in an embodiment of this application. The method includes the following steps:

[0138] S510, the monitoring agent module 401 obtains multi-dimensional indicator data of the node at the target time.

[0139] The multidimensional metrics data for a node include resource utilization metrics, load change metrics, predicted SLA default metrics, and electricity price metrics. The electricity price metrics include the regional electricity price of the area where the node is located, used to reflect dynamic costs. The resource utilization metrics include, but are not limited to, the node's GPU utilization, CPU utilization, memory utilization, and bandwidth utilization.

[0140] The load change metrics for a node can be the node load surge rate or the magnitude of load increase within a preset time period. The preset time period is a duration set in advance as needed, for example, a preset time period of 30 seconds.

[0141] The load surge rate of a node can be the increase in the node's load over the past 30 seconds. The specific calculation formula is shown in formula (2).

[0142] (2)

[0143] in, This represents the average load over the past 30 seconds. This represents the average load over the past 60 seconds. This refers to the load surge rate of the node. By using the load change metric of the node, overheating trends can be detected in advance.

[0144] The predicted SLA default metric can be the probability of SLA default within a preset future time period. This metric is used to determine whether migration operations will lead to a decline in service quality. In one specific implementation, a Long Short-Term Memory (LSTM) network model can be used as the predictive microservice. By processing node request latency, error rate, and traffic, the probability of node SLA default in the future time period can be predicted.

[0145] In cross-cloud computing pools, node resource utilization metrics only reflect the current resource usage of a node and cannot reflect its load stability, making it impossible to accurately determine a node's hot or cold status using only these metrics. A node's hot status indicates a high heat value, while a node's cold status indicates a low heat value.

[0146] For example, if a node's CPU utilization is 60% and its load consists of executing a lightweight scheduled task every 5 minutes, this node is considered a "cold" node. If the load is a high-concurrency e-commerce transaction service, this node is considered a "hot" node. Therefore, it is difficult to accurately determine a node's "cold" or "hot" status solely based on resource utilization metrics. To address this, this application introduces multi-dimensional metrics, including node resource utilization metrics, load change metrics, predicted SLA default metrics, and electricity price metrics, to determine the fluctuation of node load, whether task migration will lead to a decline in service quality, and whether the node's resource consumption is driven by cost factors, thereby accurately determining the node's "cold" or "hot" status.

[0147] It should be noted that, in this embodiment, to ensure that indicator data from different nodes can be accurately clustered according to region and business, each indicator is labeled. The label content includes, but is not limited to, the region where the node containing the indicator is located, the resource pool, the cloud vendor, instance information, and business-related attributes. By labeling the indicators, it can be ensured that subsequent modules can accurately cluster the received indicator data according to node and business dimensions.

[0148] S520, the monitoring agent module 401 sends the multi-dimensional indicator data of the nodes at the target time to the indicator bus module 402. Correspondingly, the indicator bus module 402 receives the multi-dimensional indicator data of each node at the target time.

[0149] S530, the indicator bus module 402 arranges the multi-dimensional indicator data of each node according to the time sequence to ensure the integrity of the time sequence.

[0150] S540, the indicator bus module 402 sends the sorted indicator data to the heat analysis unit 201. Correspondingly, the heat analysis unit 201 obtains the sorted indicator data.

[0151] S550, Heat Analyzer 201 obtains the multidimensional index sequence of each node within the scrolling window.

[0152] In this embodiment, the duration of the scrolling window is a fixed duration, for example, 600 seconds, and the multidimensional index sequence... It only includes multidimensional indicator data for a recent fixed time period. It can be understood that the heat analyzer 201 obtains the multidimensional indicator data of a node at time t. , the data Insert the multidimensional indicator sequence for this node within the scrolling window, while simultaneously deleting the earliest multidimensional indicator data from the sequence. That is, the multidimensional indicator sequence... It updates in real time and always retains only the latest indicator sampling data of a fixed duration, which provides stable time dimension data for subsequent data processing.

[0153] S560, the heat analysis unit 201 processes the multi-dimensional index sequence of each node to obtain the instantaneous heat of each node.

[0154] The following explanation uses a single node as an example. This node can be any one of multiple nodes.

[0155] Heat Analyzer 201 can analyze the multidimensional index sequence of nodes. Determine the dynamic quantiles. For example, the heat map analyzer 201 can determine the dynamic quantiles based on a multidimensional index sequence. Determine the 1st percentile and 99th percentile for each indicator. The 1st percentile of an indicator is the percentage of data points less than or equal to 1% of the data points for that indicator. The 99th percentile of an indicator is the percentage of data points less than or equal to 99% of the data points for that indicator.

[0156] In some examples, to mask occasional spikes, the heatmap analyzer 201 can also analyze multidimensional index sequences. The multidimensional indicator values ​​within the sequence are cropped to the 1st to 99th percentile range. That is, for the multidimensional indicator sequence... The k-th indicator can be pruned using dynamic 1st and 99th quantiles, retaining only the indicators within the 1st to 99th quantile range. For example, consider the indicator data for the k-th indicator at time t. We can use formula (3) to cut, and obtain the following cut values:

[0157] (3)

[0158] in, Let be the clipped value corresponding to the index data of the k-th index at time t. The dynamic 1% quantile of the k-th indicator. is the dynamic 99th percentile of the k-th indicator.

[0159] Then, the heat analysis unit 201 normalizes the clipped values ​​to obtain dimensionless data within a preset range. In one specific implementation, the heat analysis unit 201 can perform sliding normalization according to formula (4) to obtain the index data of the k-th index at time t. Corresponding dimensionless data :

[0160] (4)

[0161] in, For smoothing parameters, For data in the range [0,1], for example . Multidimensional indicator sequence The minimum value of the k-th indicator. Multidimensional indicator sequence The maximum value of the k-th indicator. Due to the multidimensional indicator sequence. As time progresses, both the extreme values ​​and the scaling ratio will automatically adjust to the load offset and will not fail due to long-term baseline shifts.

[0162] Heat Analyzer 201 can analyze multidimensional index sequences The corresponding dimensionless data is used to obtain the instantaneous popularity. In the specific implementation, the popularity analyzer 201 assigns weights to each multidimensional indicator, multiplies the dimensionless data corresponding to each indicator with the corresponding weight, and linearly superimposes the multiplication results of the multidimensional indicators at the same time to obtain the instantaneous popularity at that time.

[0163] For example, the heat analysis 201 uses five metrics: CPU utilization, GPU utilization, load increase over the past 30 seconds, SLA default probability for the next 60 seconds, and real-time electricity price in the node's region. Each of these metrics is assigned a weight w. The weights for CPU utilization (0.25), GPU utilization (0.25), load increase over the past 30 seconds (0.2), SLA default probability over the next 60 seconds (0.2), and the real-time electricity price of the node's region (0.1) are all considered. The heatmap analyzer 201 multiplies the data of these five indicators by their corresponding weights. For the same moment, the results of these multiplications are linearly summed to obtain the instantaneous heat of the node at that moment. For example, at time t, the instantaneous heat of the node at time t is obtained by linearly summing the results of the multiplications of these five indicators. As shown in formula (5):

[0164] (5)

[0165] It should be noted that, in this embodiment, the weight values ​​can be fine-tuned as needed. For example, they can be automatically adjusted using built-in preset rules, or they can be fine-tuned by writing them to the configuration center via hot update; this embodiment is not limited to these methods.

[0166] S560, Heat Analyzer 201 maps the instantaneous heat of each node to the heat risk coefficient of each node.

[0167] In this embodiment of the application, the heat risk coefficient of a node specifically includes the heat risk coefficients of different business levels within that node.

[0168] In this embodiment, the heat analysis unit 201 maps the instantaneous heat of a node into a heat risk coefficient that varies with the business level, ensuring that critical services have a longer buffer under high pressure, while other services give way when peak times arrive. Specifically, the heat analysis unit 201 uses the pre-acquired exponential memory-logistic curve corresponding to each business level to obtain the heat risk coefficient of different business levels in the node, achieving the goal of different business levels responding to the load according to their own characteristics.

[0169] Appendix Figure 6 A flowchart illustrating a method for obtaining the heat risk coefficient of a node based on an exponential memory-logistic curve, provided in this application embodiment, includes the following steps:

[0170] S610: Employs an exponentially weighted moving average algorithm to accumulate historical heat over the time domain, thereby obtaining the exponential memory value of each business level in the node.

[0171] Specifically, the heat analysis 201 analyzes any node and any service level of that node. This means that the service level P belongs to one of the following: critical service level (i.e., first service level), major service level (i.e., second service level), or minor service level (i.e., third service level). Define the memory constant for this service level P. and attenuation coefficient Among them, the attenuation coefficient . It is the time interval for the 401 sampling of the monitoring agent module, for example. It lasts for 2 seconds.

[0172] It should be noted that critical business processes are more critical than major business processes, and major business processes are more critical than minor business processes.

[0173] Heat Analyzer 201 combined with memory constant and attenuation coefficient According to formula (6), the instantaneous popularity and historical popularity of business level P are accumulated to obtain the exponential memory value of business level P. Formula (5) is as follows:

[0174] (6)

[0175] in, This is the exponential memory value at time t preceding time t. At the initial time t = 0, its initial memory exponential value is... . The closer to 1, the higher the exponential memory value. The greater the proportion of historical popularity, the more significant the impact.

[0176] S620: The Logistic curve is used to map the exponential memory value of each business in the node to the interval [0, 1] to obtain the heat risk coefficient of each business in the node.

[0177] In its implementation, the heat analysis unit 201 stores the business level P-index value according to the Logistic curve. Mapping to the interval [0, 1], the specific mapping relationship is shown in formula (7):

[0178] (7)

[0179] The heat risk coefficient for business level P. The critical threshold for the memory value of business level P. This is the slope parameter of the Logistic curve.

[0180] In the embodiments of this application, if , and satisfy , , ,and satisfy The heat risk coefficient curves corresponding to the three business levels set by business level P are as follows: Figure 7 As shown. Figure 7 The diagram shows the heat risk coefficient curves corresponding to the three business levels defined in the business settings. The horizontal axis of the curve represents the exponential memory value, and the vertical axis represents the heat risk coefficient. The risk coefficient for Critical business categories is lower than that for Major business categories, while the risk coefficient for Major business categories is higher than that for Minor business categories.

[0181] To enable those skilled in the art to better understand, the following description is provided in conjunction with specific embodiments.

[0182] Table 2 is a table of parameter values ​​for different service levels corresponding to a service level P provided in the embodiments of this application.

[0183]

[0184] Table 2 provides the corresponding information for the three service levels. , and .in, , , In this application embodiment, for Critical services, a large [scale / size] is used. This indicates that only when a node accumulates heat over a long period of time will the Critical service be considered a candidate for migration. Furthermore, this service utilizes a relatively large... and smaller This makes it more sensitive to historical trends, and instantaneous spikes are downplayed, which can further reduce the probability of false migrations, thereby reducing the SLA intensity of critical services. The opposite is true for non-critical services. For example, for minor services, a larger... The instantaneous spikes will not be faded, but a smaller one will be used. Any peak will quickly The value is increased to 1, which makes scheduler 202 prioritize relocating this type of task.

[0185] For example, if instantaneous heat Previous round Both are 0.30. For example, we can obtain the current round's... .

[0186] Table 3 shows the different service levels.

[0187]

[0188] Table 3 provides the corresponding service levels at time t. The value. The smaller, The smaller the value, the faster the historical decay rate. In this embodiment of the application, the Minor service... maximum.

[0189] Heat Analyzer 201 will analyze different business levels Mapped to the heat risk coefficient in the interval [0, 1] .

[0190] Table 4 shows the heat risk coefficients for different business levels.

[0191]

[0192] Table 4 provides information on different service levels. Mapped to the heat risk coefficient in the interval [0, 1] The specific mapping results. Specifically, for the same instantaneous heat and the same initial memory conditions, , , Due to differences in the business load, the heat risk coefficients for the three business levels also differ, specifically: Minor business heat risk coefficient > Major business heat risk coefficient > Critical business heat risk coefficient. In this embodiment, the scheduler 202 will migrate tasks based on the heat risk coefficient of the business in the node, prioritizing the migration of businesses with higher heat coefficients, i.e., prioritizing the migration of Minor businesses. If the node still has load pressure, it can consider migrating businesses with lower heat coefficients, i.e., Major businesses.

[0193] This application combines short-term heat with exponential memory values ​​and uses an adjustable Logistic function to map the same physical heat, creating a clear gradient among three different levels of business: Minor tasks are most sensitive to peaks and migrate first; Major tasks are of moderate sensitivity; Critical tasks only enter the migration list after prolonged heat accumulation and low-priority workloads have been relocated. This allows for rapid release of computing power when resources are scarce while preventing critical businesses from being mistakenly migrated due to momentary fluctuations.

[0194] S570, the heat analysis 201 sends the heat risk coefficient of each node to the scheduler 202. Correspondingly, the scheduler 202 receives the heat risk coefficient of each node.

[0195] In this embodiment of the application, the heat risk coefficient of a node includes the heat risk coefficients of different services within the node.

[0196] S580, scheduler 202 determines the source node and candidate node set based on the heat risk coefficient of each node.

[0197] After scheduler 202 obtains the heat risk coefficient of a node, it performs a pre-filter using a lightweight threshold. If the highest heat risk coefficient of the source node exceeds the overheat threshold, it retrieves all nodes with the lowest heat risk coefficient below the cold threshold from the global node table as candidate nodes. These candidate nodes are the destination nodes for potential load migration, thus narrowing the search space. The global node table stores all nodes across the computing pool.

[0198] In one example, scheduler 202 can determine the source node and candidate nodes through the following steps:

[0199] Step A1: Scheduler 202 scans the heat risk coefficient of each node at a fixed period to obtain the highest and lowest heat risk coefficients of each node.

[0200] Specifically, the highest and lowest heat risk coefficients of node n can be obtained by referring to formula (8):

[0201] (8)

[0202] in, The highest heat risk coefficient representing node n. is the minimum heat risk coefficient for node n.

[0203] The fixed period can be adjusted as needed, for example, a fixed period of 2 seconds.

[0204] Step A2: Determine the source node and candidate nodes.

[0205] Specifically, scheduler 202 determines whether the node's highest heat risk coefficient is greater than or equal to the overheating threshold. If so, then mark that node as the source node. For example, determine the highest heat risk coefficient for node n. Does it meet the requirements? If so, then mark node n as the source node.

[0206] Scheduler 202 queries the global node table for all nodes that meet the minimum heat risk coefficient less than or equal to the cold threshold. Nodes that meet the criteria are selected as candidate nodes. For example, all nodes that satisfy the criteria are... Node m is selected as a candidate node, and a candidate node set is obtained. Specifically, as shown in formula (9):

[0207] (9)

[0208] It should be noted that, to ensure the migration task succeeds, nodes that are unreachable due to network issues or have mismatched resource types need to be excluded. This refers to the set of candidate nodes. This excludes nodes that are unreachable from the network or whose resource types do not match.

[0209] It should be noted that for the same node, after executing a migration task, a preset cooldown period is required before triggering the migration operation again to avoid jitter. The preset cooldown period can be adjusted as needed; for example, the preset cooldown period is 60 seconds.

[0210] S590, the scheduler 202 determines whether to trigger migration in a future preset time period based on the source node and candidate node set and through the migration triggering model.

[0211] In this embodiment, for source node n, scheduler 202 first constructs a coefficient sequence consisting of the highest heat risk coefficients of source node n within a fixed duration before time t. For example, the fixed duration is 60 seconds, and the sampling time interval is... For example, the coefficient sequence corresponding to the source node n is shown in formula (10):

[0212] (10)

[0213] Wherein, the coefficient sequence This includes 30 time-varying maximum heat risk coefficients. Taking the first-order difference, we obtain the first-order difference sequence. It can capture the risk change rate characteristics of the highest heat risk coefficient of source node n over the past 60 seconds.

[0214] Scheduler 202 will use the coefficient sequence corresponding to source node n The first difference corresponding to the coefficient sequence The 32-unit GRU is input with its respective encoding to obtain the first and second encoding vectors. The scheduler 202 then concatenates the first and second encoding vectors, the mean minimum risk of each candidate node, the number of candidates, and the median link bandwidth among the candidate nodes to obtain the state vector. .

[0215] Scheduler 202 uses the trained transfer-triggered model to process the state vector The process is performed to obtain the state vector. Migration probability that triggers migration If the migration probability Greater than or equal to a probability threshold, for example Output the first flag The specific implementation of this migration triggering model is shown in formula (11):

[0216] (11)

[0217] in, The decision parameters corresponding to the migration trigger model.

[0218] If the first flag is output Scheduler 202 determines to trigger migration. If the migration probability of source node n is less than or equal to the probability threshold, the second flag Defer is output, and scheduler 202 determines to pause triggering migration.

[0219] Furthermore, to ensure robustness, migration will be directly triggered if n1 or more SLA defaults are accumulated within a preset cumulative period. If the scheduler 202 detects that the cross-computing pool is in a high-priority service large-scale expansion window, migration will be temporarily suspended to avoid link congestion.

[0220] It should be noted that n1 is a positive integer. The large-scale expansion window for high-priority services is a specific period during which critical-level or major-level services within the cross-cloud computing power pool concentrate on resource expansion. During this period, priority should be given to ensuring the stability of the resources and links required for expansion, and unnecessary load migration should be postponed to avoid resource conflicts or link congestion.

[0221] In this embodiment of the application, if the output flag is the first flag If the condition is met, the process proceeds to S5100; otherwise, scheduler 202 terminates and proceeds to the next sampling cycle.

[0222] S5100, if a migration is triggered, the scheduler 202 determines the target node and target switching time slot for the future period based on the candidate node's lowest heat risk coefficient, link bandwidth, and real-time electricity price.

[0223] In this embodiment, the scheduler 202 first obtains the characteristics of the candidate nodes. The characteristics of each candidate node include, but are not limited to, the node's link bandwidth and real-time electricity price.

[0224] In one specific implementation, the scheduler 202 first determines the set of candidate nodes. Is it an empty set? ,Right now If the condition is not met, it indicates resource scarcity and requires expansion. If the condition is met, calculate the link bandwidth for each candidate node m. And the instantaneous electricity price E(m). It should be noted that the candidate node m mentioned in this application is any candidate node in the candidate node set. Wherein, the candidate node set... This refers to the node to which the load of source node n may migrate.

[0225] In this embodiment of the application, the scheduler 202 determines the target flow cutting time slot and the target node based on the source node, the set of candidate nodes and the characteristics of each candidate node.

[0226] In this embodiment, the scheduler 202 uses a two-layer graph attention network to process the data and obtain the target flow-cutting time slot and the target node. The following is a detailed explanation in conjunction with the appendix. Figure 8 Please provide an explanation.

[0227] See Figure 8 The diagram shown is a schematic representation of a method for obtaining a target flow-cutting time slot and a target node according to an embodiment of this application. The method specifically includes the following steps:

[0228] S810: Obtain candidate image sets.

[0229] In this embodiment of the application, the scheduler 202 combines the source node n with the candidate node set. The mapping is a weighted directed graph with attributes, i.e., a candidate graph set, to facilitate the subsequent simultaneous capture of the interaction between the hot and cold status of nodes, link capacity, and real-time electricity price.

[0230] The candidate graph set includes vertices, vertex attributes, edges, and edge attributes. Vertices include the first vertex corresponding to the source node and the second vertex corresponding to the candidate node. Vertex attributes include the heat value, instantaneous electricity price, and available processor capacity of the candidate node corresponding to the second vertex at the target time. Edges include unidirectional edges from the first vertex to the second vertex. Edge attributes include the link bandwidth and round-trip latency from the source node to the candidate node. Vertex attributes describe whether a candidate node is suitable for undertaking the migration task. Edge attributes describe the reliability of migrating the task from the overheated source node to the candidate node.

[0231] In one specific implementation, the three-dimensional feature vector of each candidate node m in the candidate node set is obtained. The three-dimensional feature vector of candidate node m includes the lowest heat risk coefficient of candidate node m at the current time, the regional electricity price, and the available graphics processing unit (GPU) capacity, as shown in formula (12):

[0232] (12)

[0233] The three-dimensional feature vector of candidate node m is .in, The minimum heat risk coefficient of candidate node m at time t is represented by E, where E is the regional electricity price of candidate node m at time t. Let m be the available GPU capacity for candidate node m. To improve the accuracy of subsequent processing, this... Specifically, it can be a number that maps the available GPU capacity to the [0,1] interval. In this embodiment, the normalized 3D feature vector of the candidate node can be written into the vertex attribute table. It should be noted that when the vertex attribute table is updated, the vertex attributes in the candidate graph set are updated synchronously.

[0234] Edge properties of the edge (n, m) from source node n to candidate node m As shown in formula (13):

[0235] (13)

[0236] in, The link bandwidth from source node n to candidate node m , The round-trip time (in milliseconds) for transmitting reference information between source node n and candidate node m. Higher link bandwidth results in better round-trip time, better network transmission quality, and higher task migration efficiency.

[0237] In this embodiment of the application, to unify the units of measurement and improve the accuracy of the calculation, the link bandwidth and round-trip delay of the candidate nodes can be normalized and mapped to numbers in the range [0,1], respectively. Specifically, Scale linearly according to [0,1], taking the reciprocal of the round-trip delay and then scaling according to [0,1].

[0238] In this way, the minimum heat risk coefficient, available GPU capacity, and instantaneous electricity price of a node, as well as the link bandwidth and round-trip latency of candidate nodes, can be uniformly encapsulated in the candidate graph atlas. middle.

[0239] It should be noted that, in this embodiment of the application, all edge attributes can be written into the sparse weight matrix to obtain a |C|×1 matrix, which can be used by the graph attention network to perform graph attention calculation.

[0240] S820 divides the future preset time period into the cut-off time slot and the discrete encoding of the cut-off time slot.

[0241] For example, the future preset time period is divided into four equal parts according to the time period, resulting in four streaming time slots, which are as follows: A two-dimensional sinusoidal code is introduced for each flow-cutting time slot to obtain the discrete code corresponding to the flow-cutting time slot. For example, the discrete code for the f-th flow-cutting time slot. As shown in formula (14):

[0242] (14)

[0243] Where f is a positive integer in the range [1,4].

[0244] It should be noted that S810 and S820 can be executed simultaneously, or S810 can be executed first and then S820, or S820 can be executed first and then S810. This application does not limit this.

[0245] S830: Input the candidate graph set, the flow cut time slot, and the source node identifier into the graph attention network to obtain the score of each candidate node.

[0246] The source node identifier is used to uniquely identify the source node. For example, the source node identifier of source node n can be "source node n".

[0247] Scheduler 202 can invoke a two-layer graph attention network to process the candidate graph set, source nodes, and discrete encoding of the flow-cutting time slots to obtain scores for the candidate nodes. The following example, using source node n and candidate node m, illustrates the score obtained by the graph attention network for candidate node n.

[0248] For the h-th attention head in the l-th layer, its feature transformation parameters are set as follows: And attention weight parameters are The attention coefficient for the edge (n, m) from source node n to node m. As shown in formula (15):

[0249] (15)

[0250] in, Let be the three-dimensional feature vector of candidate node n. Let be the three-dimensional feature vector of the source node n. Let n be the edge attributes from the source node n to the candidate node m.

[0251] Attention coefficients of edge (n, m) Normalization is performed to obtain the normalized attention coefficients. As stated in formula (16):

[0252] (16)

[0253] For candidate node m at layer l, multi-head aggregation embedding As in formula (17):

[0254] (17)

[0255] Where H represents the number of attention heads in each layer of the graph attention network, e.g., H = 4. It is an Exponential Linear Unit (ELU). The four-head output splicing method is shown in formula (18):

[0256] (18)

[0257] The result obtained after concatenating the H-head outputs is input into another outer layer of the graph attention network, and processed in the same way as the l-th layer, finally obtaining the feature vector of the candidate node m. .

[0258] The graph attention network inputs the feature vector of candidate node m along with all tangent time slots into a multilayer shared perceptron to obtain a score for candidate node m. The number of scores for candidate node m corresponds one-to-one with the number of flow-cutting time slots. For example, if there are 4 flow-cutting time slots, then the score for candidate node m is 4, as shown in formula (19).

[0259] (19)

[0260] in, For learnable parameter vectors, dimensions and concatenated features Consistency, used for splicing features Weighting is applied, where b is a bias term used to adjust the scoring baseline and to adjust the offset of the results after feature weighting.

[0261] In one specific implementation, the multilayer shared perceptron includes two fully connected layers. The first fully connected layer is used to output a 32-dimensional feature vector, and the second fully connected layer is used to output a one-dimensional feature. In this multilayer perceptron, the activation function used is the rectified linear unit (ReLU).

[0262] S840: Determine the target node and target flow switching time slot based on the scores of each candidate node.

[0263] Next, we continue with the source node n and the candidate graph set. Candidate Node Set Taking a scenario with four flow-cutting time slots as an example, this explains how to determine the target node and the target flow-cutting time slot.

[0264] Scheduler 202 first obtains the set of all actions. As shown in formula (20):

[0265] (20)

[0266] The temperature-based soft-max formula is shown in formula (21):

[0267] (twenty one)

[0268] in, It is a temperature coefficient and belongs to the category of numbers greater than 0 and less than 1. Candidate image set Below, the migration-triggered model outputs the load during the time period. The probability of migrating to candidate node m. It is the set of all available actions within the same time period.

[0269] In this embodiment of the application, during the training phase, the data is taken... =1, press the action directly. Sampling is performed to ensure unbiased policy gradient estimation; temperature coefficients are included during the inference phase. Adjust the values ​​to be below 1 and above 0, for example, to 0.3 to concentrate the distribution. From all feasible action sets, select the combination with the highest probability as the target migration scheme, as shown in formula (22):

[0270] (twenty two)

[0271] in, For the target migration plan, For the target node, The target flow cut time slot.

[0272] For example, continuing to target Figure 1 In the cross-cloud computing pool scenario shown, if the source node n is private cloud 1, the candidate node set C(n,t) = {edge node 1, public cloud, edge node 2}, the preset time period is 5 minutes, and the corresponding flow switching time slot is 4, including... (0-1.25min) (1.25-2.5 min) (2.5-3.75 min) and (3.75-5 min), temperature coefficient The value obtained is 0.3. As shown in Table 4:

[0273] Table 5: Scoring Results of Candidate Nodes

[0274]

[0275] Table 5 shows the scores of all candidate nodes in different flow-cutting time slots. Different candidate nodes have different scores in different flow-cutting time slots. This reflects the scores of all action sets. ={(edge ​​node 1, (edge ​​node 1, ) (edge ​​node 1, ) (edge ​​node 1, ) (Public cloud) ), (Public cloud, ), (Public cloud, ), (Public cloud, (edge ​​node 2, (edge ​​node 2,) (edge ​​node 2,) (edge ​​node 2,) The scores corresponding to each action in the formula are given in Table 6. Substituting the scores of each action into formula (22), the probabilities of all actions can be obtained.

[0276] Table 6 shows the probability of all actions obtained.

[0277]

[0278] The migration scheme corresponding to the highest probability among all actions in the graph is taken as the target migration scheme. As shown in Table 5, the highest probability is 0.277, that is, edge node 1 is taken as the target node, T3 is the target flow cut time slot, and the target migration scheme is obtained.

[0279] S5110, scheduler 202 sends the target node and target flow switching time slot to orchestrator 203. Correspondingly, orchestrator 203 receives the target node and target flow switching time slot.

[0280] In the specific implementation, to avoid passing unexecutable target migration schemes to the execution layer, the real-time resources of the target node are verified immediately after the action is generated. Link constraints If the following conditions are met:

[0281]

[0282] This confirms that the data can be sent to orchestrator 203. Among other things, The preset resource threshold can be adjusted as needed, for example... , The preset bandwidth threshold can be adjusted as needed, for example... .

[0283] If any condition is not met, then ( If the probability of ) is set to zero, it is normalized again according to the remaining probability and resampled; if the set of possible actions is cleared, the "no suitable node" event is output, and the migration orchestrator 203 is transferred to the expansion process.

[0284] Once it is determined that transmission is possible, scheduler 202 will assign the identifier corresponding to source node n, Write them together to the migration log and send the migration log to orchestrator 203.

[0285] S5120, orchestrator 203 performs load migration based on the target node, target cut-off time slot and source node.

[0286] In this embodiment, the experience database continuously receives load migration execution results (including migration success or failure), actual time consumption, link bandwidth, electricity price, and load migration reward value. The scheduler can retrieve these records from the experience database to update the decision parameters of the migration triggering model, thereby improving the accuracy of the migration triggering model.

[0287] For the specific execution method, please refer to step S40, which will not be discussed further here.

[0288] The load migration method provided in this application integrates historical heat accumulation and instantaneous spikes through exponential memory, combined with Logistic function mapping and business priority-specific parameters, such as the memory constant. and attenuation coefficient This approach helps differentiate business priorities during migration, avoiding the problems of critical businesses being mistakenly migrated or low-priority businesses being migrated late. This helps reduce the risk of critical business SLA defaults and allows for rapid resource reallocation to handle tidal loads. Furthermore, the scheduling process is only initiated on the source node whose highest risk coefficient reaches the overheating threshold. The lowest heat risk coefficient, less than or equal to the cold threshold, is obtained from the global node table of the cross-cloud computing pool. Then, through learning decisions, the appropriate migration and its location are determined. This method reduces unnecessary computation and helps reduce the probability of network jitter. The previous stage's reinforcement learning framework, which uses a migration triggering model to determine the migration timing and a graph attention network to determine the target node and target flow switching time slot, enables rapid response in the decision-making process and accurate determination of the migration plan.

[0289] It should be noted that the core technical solutions in the embodiments of this application are illustrated using a single source node load migration scenario as an example, but its technical architecture naturally supports the expansion requirements of parallel migration of multiple source nodes. For scenarios in practical applications where there are multiple source nodes (such as multiple overheated nodes appearing simultaneously in a cross-cloud computing power pool) and all of them need to perform load migration, the load migration method provided in this application can be directly reused for each source node without the need for additional complex adaptation logic.

[0290] Specifically, the migration scheduling of multiple source nodes can be achieved as follows: At the target time, the source nodes that need to be migrated and the candidate nodes to which the load may be migrated are determined from multiple nodes in the cross-cloud computing pool; based on the heat value sequence of each source node at the target time, the heat value sequence of heat value changes, the heat value of the candidate node corresponding to each source node at the target time, and the link bandwidth from the source node to the candidate node, it is determined whether to trigger the migration of each source node in a preset future period. If at least two source nodes among multiple source nodes are triggered to migrate, the target nodes corresponding to each source node to be migrated can be determined from the candidate nodes based on the heat value, real-time electricity price, and link bandwidth of the candidate nodes corresponding to each source node at the target time, and the workload is migrated from the source node to the target node corresponding to the source node.

[0291] This extended implementation does not require complex logical reconstruction by those skilled in the art. It only requires parallel scheduling of tasks and coordination of resource conflicts based on existing modules to quickly adapt to the actual scenario of simultaneous migration of multiple source nodes, fully ensuring the global self-balancing capability and operational reliability of the cross-cloud computing power pool under large-scale tidal loads.

[0292] against Figures 2-8 In addition to the load migration method shown, this application also provides a load migration apparatus for implementing the above-described load migration method.

[0293] See Figure 9 The diagram shown is a structural schematic of a load migration device provided in an embodiment of this application. Figure 9 As shown, the device 900 includes:

[0294] The first determining unit 901 is used to determine the source node and candidate node of the plurality of nodes at a target time; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated.

[0295] The second determining unit 902 is used to determine whether to trigger a load migration task in a future preset period based on the heat value sequence and heat value change sequence of the source node at the target time, and the heat value and link bandwidth of the candidate node at the target time.

[0296] Wherein, the link bandwidth of the candidate node represents the link bandwidth from the source node to the candidate node, the future preset time period is the time period after the target time; the heat value sequence is a sequence composed of the heat values ​​of the source node within the target time period, the target time period is the time period before the target time; the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence;

[0297] The third determining unit 903 is used to determine a load migration strategy based on the heat value, real-time electricity price and link bandwidth of the candidate node at the target time if a load migration task is triggered, so as to perform load migration according to the load migration strategy; the load migration strategy includes a target node, and the target node is one of the candidate nodes.

[0298] Optionally, determining whether to trigger a load migration task in a future preset time period based on the heat value sequence and heat value change sequence of the source node at the target time, and the heat value and link bandwidth of the candidate node at the target time, includes:

[0299] Obtain the first encoding vector corresponding to the heat value sequence, and the second encoding vector corresponding to the heat value change sequence;

[0300] The first encoding vector, the second encoding vector, the heat value of the candidate node at the target time, and the link bandwidth are concatenated to obtain the state vector.

[0301] The state vector is processed using a migration triggering model to determine whether a load migration task will be triggered within a preset future time period. The input of the migration triggering model is the state vector, and the output of the migration triggering model is a Boolean flag. When the Boolean flag is a first flag, the load migration task is determined to be triggered. When the Boolean flag is a second flag, the load migration task is paused.

[0302] Optionally, the transfer triggering model includes a decision network and a value network; the transfer triggering model is a trained model, and during the training process, the transfer triggering model adjusts the decision parameters of the decision network according to the state value obtained by the value network;

[0303] The step of processing the state vector using a migration triggering model to determine whether a load migration task will be triggered within the preset future time period includes:

[0304] The state vector is processed by the trained decision network to obtain the transition probability;

[0305] If the migration probability is greater than or equal to the probability threshold, output the first flag; if the migration probability is less than the probability threshold, output the second flag.

[0306] Optionally, the migration-triggered model is trained using a proximal policy optimization algorithm, and an immediate reward function is used during the training process;

[0307] The instant reward function is negatively correlated with the reward value of migration costs and positively correlated with the improvement in the achievement rate of Service Level Agreements (SLAs) within the preset future time period.

[0308] Optionally, determining the source node and candidate nodes of the plurality of nodes includes:

[0309] Obtain the popularity value of each of the plurality of nodes at the target time; the popularity value of a node includes the popularity value corresponding to each service in the node;

[0310] Based on the heat value of each node at the target time, source nodes and candidate nodes are determined from the plurality of nodes; the highest heat value of the source node at the target time is greater than or equal to an overheating threshold, the lowest heat value of the candidate node at the target time is less than or equal to a slightly cold threshold, and the overheating threshold is greater than the slightly cold threshold; the highest heat value of the source node is the maximum value of the heat value corresponding to each service in the source node, and the lowest heat value of the candidate node is the minimum value of the heat value corresponding to each service in the candidate node.

[0311] Optionally, determining the load migration strategy based on the candidate node's popularity value, real-time electricity price, and link bandwidth at the target time includes:

[0312] Using a graph attention network, the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time are processed to determine the score of the candidate nodes; the score of the candidate nodes is used to describe the relative priority of migrating load to the candidate nodes;

[0313] The load migration strategy is determined based on the scores of the candidate nodes.

[0314] Optionally, the load migration strategy further includes a target flow-switching time slot, which indicates the time slot for load migration. In the application embodiment, the preset future time period can be flow-switched according to a preset rule to obtain N flow-switching time slots; where N is an integer greater than or equal to 1.

[0315] The graph attention network is used to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes, including:

[0316] Using the graph attention network, the heat value, instantaneous electricity price, link bandwidth, and N flow switching time slots of the candidate nodes at the target time are processed to determine the scores of the N candidate nodes; the scores of the candidate nodes correspond one-to-one with the flow switching time slots.

[0317] Optionally, before processing the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time using the graph attention network to determine the score of the candidate nodes, a candidate graph set is obtained; the candidate graph set includes vertices, vertex attributes, edges, and edge attributes; the vertices include a first vertex corresponding to the source node and a second vertex corresponding to the candidate node; the vertex attributes include the popularity value, instantaneous electricity price, and available processor capacity of the candidate node corresponding to the second vertex at the target time; the edges include unidirectional edges from the first vertex to the second vertex; and the edge attributes include the link bandwidth and round-trip delay from the source node to the candidate node.

[0318] Then, the candidate graph set is processed using a graph attention network to determine the score of the candidate node.

[0319] Optionally, the device 900 further includes a verification unit for determining the real-time resources and real-time link bandwidth of the target node; if the real-time resources are greater than or equal to a preset resource threshold and the real-time link bandwidth is greater than or equal to a preset bandwidth threshold, load migration is performed according to the load migration strategy.

[0320] The load migration device provided in this application can capture the dynamic trend of heat value changes through two indicators: the heat value change trend and the heat value change rate of the source node. Compared with using a fixed threshold to trigger migration, it can improve the accuracy of triggering migration. By comprehensively considering the link bandwidth and heat value of candidate nodes, it avoids secondary overload or migration interruption after load migration, further improving the accuracy of triggering migration. By improving the accuracy of triggering migration, the resource utilization and operational reliability of cross-cloud computing pools can be improved. Furthermore, if the load migration system triggers a migration task, it can determine the target node from the candidate nodes based on the heat value, real-time electricity price, and link bandwidth of the candidate nodes at the target time, and migrate the workload to the target node. When determining where to migrate, comprehensively considering the heat value, real-time electricity price, and link bandwidth of the candidate nodes at the target time can achieve low-cost and high-reliability migration of workloads to the target node. Moreover, by decoupling whether to trigger migration and where to migrate, this method can shorten the determination time and improve scheduling efficiency compared to directly determining where to migrate and whether to trigger migration in a single stage.

[0321] According to the method provided in the embodiments of this application, this application also provides a chip system, which includes one or more processors for calling and executing instructions stored in memory, thereby causing the method described in the embodiments of this application to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0322] The chip system may include input circuits or interfaces for transmitting information or data, and output circuits or interfaces for receiving information or data.

[0323] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0324] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0325] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or it may include both volatile memory and non-volatile memory.

[0326] In the embodiments of this application, the terms and English abbreviations are exemplary examples given for ease of description and should not be construed as limiting the application in any way. This application does not preclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0327] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0328] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

Claims

1. A load migration method, characterized in that, Applied to a cross-cloud computing power pool, wherein the cross-cloud computing power pool comprises multiple nodes, the method includes: At the target time, the source node and candidate node of the plurality of nodes are determined; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated. Based on the heat value sequence and heat value change sequence of the source node at the target time, as well as the heat value and link bandwidth of the candidate node at the target time, determine whether to trigger a load migration task in a future preset period of time; Wherein, the link bandwidth of the candidate node represents the link bandwidth from the source node to the candidate node, the future preset time period is the time period after the target time; the heat value sequence is a sequence composed of the heat values ​​of the source node within the target time period, the target time period is the time period before the target time; the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence; If a load migration task is triggered, a load migration strategy is determined based on the popularity value, real-time electricity price, and link bandwidth of the candidate node at the target time, and load migration is performed according to the load migration strategy; the load migration strategy includes a target node, which is one of the candidate nodes.

2. The method according to claim 1, characterized in that, The step of determining whether to trigger a load migration task in a future preset time period based on the heat value sequence and heat value change sequence of the source node at the target time, and the heat value and link bandwidth of the candidate node at the target time, includes: Obtain the first encoding vector corresponding to the heat value sequence, and the second encoding vector corresponding to the heat value change sequence; The first encoding vector, the second encoding vector, the heat value of the candidate node at the target time, and the link bandwidth are concatenated to obtain the state vector. The state vector is processed using a migration triggering model to determine whether a load migration task will be triggered within a preset future time period. The input of the migration triggering model is the state vector, and the output of the migration triggering model is a Boolean flag. When the Boolean flag is a first flag, the load migration task is determined to be triggered. When the Boolean flag is a second flag, the load migration task is paused.

3. The method according to claim 2, characterized in that, The migration triggering model includes a decision network and a value network; the migration triggering model is a trained model, and during the training process, the decision parameters of the decision network are adjusted according to the state value obtained by the value network. The step of processing the state vector using a migration triggering model to determine whether a load migration task will be triggered within the preset future time period includes: The state vector is processed by the trained decision network to obtain the transition probability; If the migration probability is greater than or equal to the probability threshold, output the first flag; if the migration probability is less than the probability threshold, output the second flag.

4. The method according to claim 3, characterized in that, The migration-triggered model is trained using a proximal policy optimization algorithm, and an immediate reward function is used during the training process. The instant reward function is negatively correlated with the reward value of migration costs and positively correlated with the improvement in the achievement rate of Service Level Agreements (SLAs) within the preset future time period.

5. The method according to claim 1, characterized in that, Determining the source node and candidate nodes of the plurality of nodes includes: Obtain the popularity value of each of the plurality of nodes at the target time; the popularity value of a node includes the popularity value corresponding to each service in the node; Based on the heat value of each node at the target time, source nodes and candidate nodes are determined from the plurality of nodes; the highest heat value of the source node at the target time is greater than or equal to an overheating threshold, the lowest heat value of the candidate node at the target time is less than or equal to a slightly cold threshold, and the overheating threshold is greater than the slightly cold threshold; the highest heat value of the source node is the maximum value of the heat value corresponding to each service in the source node, and the lowest heat value of the candidate node is the minimum value of the heat value corresponding to each service in the candidate node.

6. The method according to claim 1, characterized in that, The process of determining the load migration strategy based on the candidate node's popularity value, real-time electricity price, and link bandwidth at the target time includes: Using a graph attention network, the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time are processed to determine the score of the candidate nodes; the score of the candidate nodes is used to describe the relative priority of migrating load to the candidate nodes; The load migration strategy is determined based on the scores of the candidate nodes.

7. The method according to claim 6, characterized in that, The load migration strategy further includes a target flow-switching time slot, the target flow-switching time slot indicating the time slot for load migration, and the method further includes: The preset future time period is cut into streams according to preset rules to obtain N stream cutting time slots; where N is an integer greater than or equal to 1. The graph attention network is used to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes, including: Using the graph attention network, the heat value, instantaneous electricity price, link bandwidth, and N flow switching time slots of the candidate nodes at the target time are processed to determine the scores of the N candidate nodes; the scores of the candidate nodes correspond one-to-one with the flow switching time slots.

8. The method according to claim 6, characterized in that, Before using a graph attention network to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes, the method includes: Obtain a candidate graph set; the candidate graph set includes vertices, vertex attributes, edges, and edge attributes; the vertices include a first vertex corresponding to the source node and a second vertex corresponding to the candidate node; the vertex attributes include the heat value, instantaneous electricity price, and available processor capacity of the candidate node corresponding to the second vertex at the target time; the edges include a one-way edge from the first vertex to the second vertex; the edge attributes include the link bandwidth and round-trip latency from the source node to the candidate node. The step of using a graph attention network to process the popularity value, instantaneous electricity price, and link bandwidth of the candidate nodes at the target time to determine the score of the candidate nodes includes: processing the candidate graph set using a graph attention network to determine the score of the candidate nodes.

9. The method according to claim 6, characterized in that, After determining the load migration strategy, the method further includes: Determine the real-time resources and real-time link bandwidth of the target node; If the real-time resources are greater than or equal to a preset resource threshold, and the real-time link bandwidth is greater than or equal to a preset bandwidth threshold, load migration is performed according to the load migration strategy.

10. A load transfer device, characterized in that, Applied to a cross-cloud computing power pool, wherein the cross-cloud computing power pool includes multiple nodes, the device includes: The first determining unit is used to determine the source node and candidate node of the plurality of nodes at a target time; the source node represents the node that needs to be migrated, and the candidate node represents the destination node where the load may be migrated. The second determining unit is used to determine whether to trigger a load migration task in a future preset period based on the heat value sequence and heat value change sequence of the source node at the target time, as well as the heat value and link bandwidth of the candidate node at the target time. Wherein, the link bandwidth of the candidate node represents the link bandwidth from the source node to the candidate node, the future preset time period is the time period after the target time; the heat value sequence is a sequence composed of the heat values ​​of the source node within the target time period, the target time period is the time period before the target time; the heat value change sequence is a sequence obtained by first-order difference of the heat value sequence; The third determining unit is used to determine a load migration strategy based on the popularity value, real-time electricity price and link bandwidth of the candidate node at the target time if a load migration task is triggered, so as to perform load migration according to the load migration strategy; the load migration strategy includes a target node, which is one of the candidate nodes.

Images