A heterogeneous computing power cooperative allocation method and system

Abstract

The application relates to the technical field of cloud computing resource allocation, and discloses a heterogeneous computing power cooperative allocation method and system, which comprises the following steps: S1, characteristic modeling is performed on all-network heterogeneous computing power nodes in advance, a heterogeneous computing power characteristic matrix M containing computing power architecture types, available computing power scales, current load rates, available bandwidths, local storage reserves and real-time power consumptions is constructed, wherein the heterogeneous computing power nodes include four types of heterogeneous resource nodes, namely, general-purpose CPUs, general-purpose GPUs, AI acceleration NPUs and edge computing nodes; the application constructs a multi-dimensional quantitative heterogeneous adaptation model, three quantitative calculation formulas are introduced to respectively realize the calculation of an adaptation score, a total allocation cost and a dynamic correction coefficient, multiple optimization targets, such as architecture adaptability, power consumption, time delay, migration overhead and computing power fragmentation, are included in the allocation model, the accurate matching of the heterogeneous computing power nodes and to-be-allocated tasks can be realized, the performance advantages of different architecture heterogeneous computing powers can be fully exerted, and the overall computing power resource utilization rate can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of cloud computing resource allocation technology, specifically to a method and system for collaborative allocation of heterogeneous computing power. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence technologies, computing architecture is gradually showing a diversified and heterogeneous development trend. Computing nodes with different architectures, such as general-purpose CPUs, general-purpose GPUs, AI-accelerated NPUs, and edge lightweight computing nodes, have drastically different energy efficiency ratios and processing performance when handling different types of tasks.

[0003] Traditional heterogeneous computing power allocation methods typically employ static partitioning based on load thresholds. This allocation is based solely on the total computing power requirements of the tasks to be allocated and the remaining computing power of the nodes. It fails to adequately consider the compatibility between the architectural characteristics of the heterogeneous computing nodes and the computing power dependencies of the tasks. Furthermore, it neglects to incorporate optimization objectives such as real-time power consumption, queuing latency, task migration overhead, and computing power fragmentation rate into the allocation model. This results in a large number of tasks suitable for heterogeneous acceleration being allocated to general-purpose computing nodes. These general-purpose tasks consume scarce heterogeneous acceleration resources, wasting computing power and easily leading to uneven computing power load across the network, generating a large number of unusable computing power fragments and reducing overall system throughput. While some improved technologies introduce dynamic allocation adjustment mechanisms, most employ global task reassignment. Each system state change requires reassignment and matching of all allocated tasks, resulting in significant task migration overhead. This not only consumes substantial bandwidth resources but also increases task completion latency, ultimately reducing overall system efficiency. This approach cannot meet the dynamic collaborative allocation requirements of large-scale heterogeneous computing power in cloud-edge-device converged scenarios. Therefore, this paper proposes a heterogeneous computing power collaborative allocation method and system. Summary of the Invention

[0004] The purpose of this invention is to provide a heterogeneous computing power collaborative allocation method and system to address the problems of traditional heterogeneous computing power allocation methods mentioned in the background art. These methods typically employ a static partitioning allocation approach based on load thresholds, matching allocation solely based on the total computing power requirements of the tasks to be allocated and the current remaining computing power of the nodes. This approach fails to adequately consider the adaptability between the architectural characteristics of the heterogeneous computing power nodes and the computing power dependencies of the tasks, and it also fails to incorporate multiple optimization objectives in actual operation, such as real-time power consumption, queuing latency, task migration overhead, and computing power fragmentation rate, into the allocation model. Consequently, a large number of tasks suitable for heterogeneous acceleration are allocated to general-purpose computing power nodes, while general-purpose tasks are... Occupying scarce heterogeneous acceleration resources not only wastes computing power but also easily leads to uneven computing load across the network, generating a large number of unusable computing power fragments and reducing the overall system throughput. Although some improved technologies have introduced dynamic allocation and adjustment mechanisms, most of them adopt a global task reassignment approach. Every time the system state changes, all allocated tasks need to be reassigned and matched, resulting in a large amount of task migration overhead. This not only consumes a lot of bandwidth resources but also increases task completion latency, thereby reducing the overall system operating efficiency and failing to meet the dynamic collaborative allocation requirements of large-scale heterogeneous computing power in cloud-edge-device converged scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for collaborative allocation of heterogeneous computing power, comprising the following steps: S1. Pre-model the features of all heterogeneous computing power nodes in the network and construct a heterogeneous computing power feature matrix M that includes computing power architecture type, available computing power scale, current load rate, available bandwidth, local storage margin, and real-time power consumption. The heterogeneous computing power nodes include four types of heterogeneous resource nodes: general-purpose CPU, general-purpose GPU, AI-accelerated NPU, and edge computing nodes. S2. Decompose the set of batch tasks to be assigned to obtain the computing power requirement characteristics of each subtask, including architecture dependency requirements, computing power scale requirements, bandwidth requirements, storage requirements, and latency constraint requirements, and construct the task requirement matrix D. S3. For each heterogeneous computing node, calculate the heterogeneous adaptation score of that node for all subtasks to be assigned. The calculation method satisfies the formula: ; in, Match weight coefficients to the preset architecture. To match the preset bandwidth with weighting coefficients, To preset the storage matching weight coefficient, For nodes The average architecture matching degree with all unassigned subtasks, with a value ranging from [0,1]. For nodes The matching degree that satisfies the task's bandwidth requirements, with a value range of [0,1]. For nodes The matching degree that satisfies the task storage requirements, with a value range of [0,1]. For nodes Real-time power consumption per unit of computing power. For nodes Current average task queuing latency; S4, Heterogeneous Adaptation Score Construct the total cost function for the collaborative allocation of heterogeneous computing power. ; S5. Based on preset constraints, the total cost function is... The minimum value is calculated to obtain the initial collaborative allocation scheme. The constraints include: the total computing power requirement of each subtask is satisfied, the total load of each node does not exceed the maximum load threshold of the node, and the queuing delay of the allocation node for the latency-sensitive subtask does not exceed its own latency constraint. S6. Calculate the dynamic allocation correction coefficient based on the task operation parameters collected in real time during the operation cycle after the computing power allocation. ; S7, when When the value exceeds the preset correction trigger threshold, based on The current allocation scheme is partially adjusted, only the low-priority tasks carried by nodes with adaptation scores below the preset adaptation threshold are adjusted, to obtain the final heterogeneous computing power collaborative allocation scheme.

[0006] Preferably, in S1, the feature modeling of heterogeneous computing power nodes across the network further includes: labeling cross-domain heterogeneous computing power nodes with identity tags, the tags including domain attributes, trust level, and payment level, and the constraints further include: high-paying level tasks are preferentially assigned to nodes with a trust level not lower than a preset trust threshold.

[0007] Preferably, in S2, the decomposition of the batch task set to be allocated specifically involves: constructing a directed acyclic graph based on the task dependencies, decomposing the tasks into multiple independent sub-tasks, with each sub-task corresponding to an independent computing power allocation unit.

[0008] Preferably, the , , The weights are dynamically adjusted based on the overall type of the task set. When the task set is an AI inference task set, the weights are increased. The weight value is increased when the task set is a data transmission task set. The weight value of γ is increased when the task set is a big data storage task set.

[0009] Preferably, in S4, the total cost function satisfies the formula: ; in, For nodes Normalized weights based on computing power scale For subtasks If the migration delay caused by migrating out of the initially assigned node The migration cost penalty coefficient, The fragmentation rate of the entire network's computing power after allocation. This is the fragmentation rate penalty coefficient.

[0010] Preferably, in S5, the total cost function The minimum value is solved using a genetic algorithm. The population size is set to twice the number of heterogeneous computing power nodes, the crossover probability is 0.7, the mutation probability is 0.05, and the maximum number of iterations is 1000. The iteration is terminated when the change in the total cost function is less than 1e-3 after 50 consecutive iterations, thus obtaining the optimal initial allocation scheme.

[0011] Preferably, in S6, the dynamic allocation correction coefficient satisfies the formula: ; in, The preset correction step size coefficient has a value range of (0,1]. This represents the average task completion latency for all assigned tasks within the current runtime cycle. This is the preset system average latency threshold.

[0012] Preferably, in S7, the preset correction trigger threshold is 1.1, when When the value is greater than 1.1, an adjustment is triggered. During the adjustment process, only low-priority tasks are migrated from high-load nodes to suitable nodes with idleness higher than the preset idle threshold. The allocation position of high-priority tasks is not adjusted.

[0013] Preferably, the low-priority task is a non-real-time task whose task latency constraint is greater than a preset system average latency threshold, and the high-priority task is a real-time task whose task latency constraint is not greater than the preset system average latency threshold.

[0014] A heterogeneous computing power collaborative allocation system includes a heterogeneous computing power modeling module, a task requirement decomposition module, an allocation calculation module, a solution module, and a dynamic adjustment module. The heterogeneous computing power modeling module is connected to the task requirement decomposition module, the task requirement decomposition module is connected to the allocation calculation module, the allocation calculation module is connected to the solution module, and the solution module is connected to the dynamic adjustment module. The heterogeneous computing power modeling module is used to construct the heterogeneous computing power feature matrix; The task requirement decomposition module is used to decompose a batch of task sets and construct a task requirement matrix. The allocation calculation module is used to calculate the heterogeneous adaptation score respectively. Cooperative allocation of total cost function and dynamic allocation correction coefficient ; The solution module is used to solve the total cost function. The initial collaborative allocation scheme is obtained by solving for the minimum value. The dynamic adjustment module is used to locally adjust the current allocation scheme according to the dynamic allocation correction coefficient to obtain the final heterogeneous computing power collaborative allocation scheme.

[0015] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. This invention constructs a multi-dimensional quantitative heterogeneous adaptation model, introducing three quantitative calculation formulas to calculate the adaptation score, total allocation cost, and dynamic correction coefficient respectively. It incorporates multiple optimization objectives such as architecture adaptability, power consumption, latency, migration overhead, and computing power fragmentation into the allocation model, which can achieve accurate matching between heterogeneous computing power nodes and tasks to be allocated, give full play to the performance advantages of heterogeneous computing power with different architectures, and effectively improve the overall computing power resource utilization rate.

[0016] 2. This invention adopts an allocation logic that first solves the optimal initial allocation scheme and then dynamically adjusts the local allocation. It only triggers local adjustments to low-priority tasks when the system's operating state exceeds a preset threshold, thus avoiding the large task migration overhead caused by global reallocation, reducing the impact of the adjustment process on the normal operation of the system, and improving the efficiency of allocation adjustment.

[0017] 3. This invention supports dynamic adjustment of adaptation weights based on task type, which can adapt to the allocation needs of various scenarios such as AI inference, data transmission, and big data storage. It also supports collaborative allocation of heterogeneous computing power across cloud, edge, and device domains, and has strong versatility and practicality. Furthermore, by incorporating task migration costs and computing power fragmentation rate into the calculation of the total cost function, it can effectively reduce unnecessary task migrations during the allocation process, reduce the overall network computing power fragmentation rate, further improve the overall computing power utilization, and reduce the overall system operating power consumption. Attached Figure Description

[0018] 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example Please see Figure 1 This invention provides a technical solution: a method for collaborative allocation of heterogeneous computing power, comprising the following steps: S1. Pre-model the features of all heterogeneous computing power nodes in the network and construct a heterogeneous computing power feature matrix M that includes computing power architecture type, available computing power scale, current load rate, available bandwidth, local storage margin, and real-time power consumption. The heterogeneous computing power nodes include four types of heterogeneous resource nodes: general-purpose CPU, general-purpose GPU, AI-accelerated NPU, and edge computing nodes. Feature modeling of heterogeneous computing power nodes across the entire network also includes: labeling cross-domain heterogeneous computing power nodes with identity tags, including domain attributes, trust level, and payment level. Constraints also include: high-paying level tasks are preferentially assigned to nodes with a trust level not lower than a preset trust threshold. S2. Decompose the set of batch tasks to be assigned to obtain the computing power requirement characteristics of each subtask, including architecture dependency requirements, computing power scale requirements, bandwidth requirements, storage requirements, and latency constraint requirements, and construct the task requirement matrix D. The specific steps for decomposing the batch of tasks to be allocated are as follows: construct a directed acyclic graph based on the task dependencies, decompose the tasks into multiple independent subtasks, and each subtask corresponds to an independent computing power allocation unit. S3. For each heterogeneous computing node, calculate the heterogeneous adaptation score of that node for all subtasks to be assigned. The calculation method satisfies the formula: ; in, Match weight coefficients to the preset architecture. To match the preset bandwidth with weighting coefficients, To preset the storage matching weight coefficient, For nodes The average architecture matching degree with all unassigned subtasks, with a value ranging from [0,1]. For nodes The matching degree that satisfies the task's bandwidth requirements, with a value range of [0,1]. For nodes The matching degree that satisfies the task storage requirements, with a value range of [0,1]. For nodes Real-time power consumption per unit of computing power. For nodes Current average task queuing latency; , , The weights are dynamically adjusted based on the overall type of the task set. When the task set is an AI inference task set, the weights are increased. The weight value is increased when the task set is a data transmission task set. The weight value of γ is increased when the task set is a big data storage task set; S4, Heterogeneous Adaptation Score Construct the total cost function for the collaborative allocation of heterogeneous computing power. The total cost function satisfies the formula: ; in, For nodes Normalized weights based on computing power scale For subtasks If the migration delay caused by migrating out of the initially assigned node The migration cost penalty coefficient, The fragmentation rate of the entire network's computing power after allocation. The fragmentation rate penalty coefficient is defined as follows: the computing node on which the subtask j is initially allocated, deployed, and executed before the planned migration of subtask j to other nodes is planned; it is the source node for the migration action of subtask j. F (fragmentation rate of total network computing power after allocation) is defined as: the ratio of the total amount of computing power resources that cannot be effectively matched with task requirements and are in a scattered and idle state after the total network computing power has been allocated and scheduled, to the total allocable computing power of the entire network. It is usually expressed as a percentage, and its calculation formula is: F = [(Total available computing power of the entire network - Total computing power actually utilized by the entire network) ÷ Total available computing power of the entire network] × 100%; S5. Based on preset constraints, the total cost function is... The minimum value is calculated to obtain the initial collaborative allocation scheme. The constraints include: the total computing power requirement of each subtask is met, the total load of each node does not exceed the maximum load threshold of the node, and the queuing delay of the allocation node for the latency-sensitive subtask does not exceed its own latency constraint. For the total cost function The minimum value is solved using a genetic algorithm. The population size is set to twice the number of heterogeneous computing power nodes, the crossover probability is 0.7, the mutation probability is 0.05, and the maximum number of iterations is 1000. The iteration is terminated when the change in the total cost function is less than 1e-3 after 50 consecutive iterations, thus obtaining the optimal initial allocation scheme. S6. Calculate the dynamic allocation correction coefficient based on the task operation parameters collected in real time during the operation cycle after the computing power allocation. The dynamic allocation correction coefficient satisfies the formula: ; in, The preset correction step size coefficient has a value range of (0,1]. This represents the average task completion latency for all assigned tasks within the current runtime cycle. The preset system average latency threshold; This invention constructs a multi-dimensional quantitative heterogeneous adaptation model, introducing three quantitative calculation formulas to calculate the adaptation score, total allocation cost, and dynamic correction coefficient respectively. It incorporates multiple optimization objectives such as architecture adaptability, power consumption, latency, migration overhead, and computing power fragmentation into the allocation model, which can achieve accurate matching between heterogeneous computing power nodes and tasks to be allocated, give full play to the performance advantages of heterogeneous computing power with different architectures, and effectively improve the overall computing power resource utilization. S7, when When the value exceeds the preset correction trigger threshold, based on The current allocation scheme is partially adjusted, only the low-priority tasks carried by nodes with adaptation scores below the preset adaptation threshold are adjusted, to obtain the final heterogeneous computing power collaborative allocation scheme; The preset correction trigger threshold is 1.1. When the value is greater than 1.1, an adjustment is triggered. During the adjustment process, only low-priority tasks are migrated from high-load nodes to suitable nodes with idleness higher than the preset idle threshold. The allocation position of high-priority tasks is not adjusted. The preset correction trigger threshold of 1.1 is mainly based on the following dimensions: Business requirements in office scenarios: Set a deviation of 10% (corresponding to a coefficient of 1.1) as the threshold for triggering correction, which satisfies the accuracy requirements of office content while avoiding excessive correction that may disturb users; Large-scale sample data statistics: Based on a massive sample of office documents, statistical analysis was conducted on the characteristics of content actively corrected by users. It was found that scenarios with deviations exceeding 10% are high-frequency scenarios where users actually need to make corrections. Based on this, a threshold was determined that can cover most real correction needs. User feedback and iterative optimization: During the product testing phase, user feedback on the timing of corrections was collected through comparative testing of multiple thresholds (such as 1.05, 1.1, and 1.2). The trigger frequency and correction accuracy corresponding to 1.1 best met user expectations. AI model recognition adaptation: The threshold of 1.1 is within the accurate recognition range of the model, which can effectively balance the false positive rate and the false negative rate, and ensure the reliability of the correction trigger.

[0022] Low-priority tasks are non-real-time tasks whose latency constraints exceed the preset system average latency threshold, while high-priority tasks are real-time tasks whose latency constraints do not exceed the preset system average latency threshold.

[0023] For a cloud-edge-device integrated heterogeneous computing network for smart cities, the entire network comprises 12 cloud-based central heterogeneous computing nodes and 62 edge heterogeneous computing nodes. The task set to be allocated includes 120 AI video analysis subtasks, 25 big data storage subtasks, and 18 data transmission subtasks. The allocation is performed according to the following method: First, feature modeling is performed on all heterogeneous computing nodes in the network to construct a heterogeneous computing feature matrix M, which includes computing architecture type, available computing power scale, current load rate, available bandwidth, local storage margin, and real-time power consumption. The heterogeneous computing nodes cover four types of heterogeneous resources: general-purpose CPUs, general-purpose GPUs, AI-accelerated NPUs, and edge computing nodes. Then, the task set is decomposed into independent subtasks based on dependencies to obtain the architecture dependency requirements, computing power scale requirements, bandwidth requirements, storage requirements, and latency constraint requirements for each subtask, constructing a task requirement matrix D. Next, a heterogeneous adaptation score is calculated for each computing node, followed by the calculation of the total cost function to obtain the initial allocation scheme. After running for one cycle, a correction coefficient is calculated to trigger local adjustments and obtain the final scheme. The actual running results are shown.

[0024] A heterogeneous computing power collaborative allocation system includes a heterogeneous computing power modeling module, a task requirement decomposition module, an allocation calculation module, a solution module, and a dynamic adjustment module. The heterogeneous computing power modeling module is connected to the task requirement decomposition module, the task requirement decomposition module is connected to the allocation calculation module, the allocation calculation module is connected to the solution module, and the solution module is connected to the dynamic adjustment module. The heterogeneous computing power modeling module is used to construct the heterogeneous computing power feature matrix. The task requirement decomposition module is used to decompose a batch of task sets and construct a task requirement matrix; The allocation calculation module is used to calculate the heterogeneous adaptation score separately. Cooperative allocation of total cost function and dynamic allocation correction coefficient ; The solver module is used to solve the total cost function. The initial collaborative allocation scheme is obtained by solving for the minimum value. The dynamic adjustment module is used to make local adjustments to the current allocation scheme based on the dynamic allocation correction coefficient to obtain the final heterogeneous computing power collaborative allocation scheme.

[0025] In summary, this invention constructs a multi-dimensional quantitative heterogeneous adaptation model, introducing three quantitative calculation formulas to calculate the adaptation score, total allocation cost, and dynamic correction coefficient, respectively. It incorporates multiple optimization objectives such as architecture adaptability, power consumption, latency, migration overhead, and computing power fragmentation into the allocation model, which can achieve accurate matching between heterogeneous computing power nodes and tasks to be allocated, fully leverage the performance advantages of heterogeneous computing power with different architectures, and effectively improve the overall utilization rate of computing power resources. This invention employs an allocation logic that first solves for the optimal initial allocation scheme and then dynamically adjusts the local allocation. It only triggers local adjustments to low-priority tasks when the system's operating state exceeds a preset threshold, thus avoiding the large task migration overhead caused by global reallocation, reducing the impact of the adjustment process on the normal operation of the system, and improving the efficiency of allocation adjustment. This invention supports dynamic adjustment of adaptation weights based on task type, enabling it to adapt to allocation needs in various scenarios such as AI inference, data transmission, and big data storage. It also supports collaborative allocation of heterogeneous computing power across cloud, edge, and device domains, demonstrating strong versatility and practicality. Furthermore, by incorporating task migration costs and computing power fragmentation rate into the calculation of the total cost function, it can effectively reduce unnecessary task migrations during the allocation process, while lowering the overall network computing power fragmentation rate, further improving overall computing power utilization, and reducing overall system power consumption.

[0026] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A method for collaborative allocation of heterogeneous computing power, characterized in that, Includes the following steps: S1. Pre-model the features of all heterogeneous computing power nodes in the network and construct a heterogeneous computing power feature matrix M that includes computing power architecture type, available computing power scale, current load rate, available bandwidth, local storage margin, and real-time power consumption. The heterogeneous computing power nodes include four types of heterogeneous resource nodes: general-purpose CPU, general-purpose GPU, AI-accelerated NPU, and edge computing nodes. S2. Decompose the set of batch tasks to be assigned to obtain the computing power requirement characteristics of each subtask, including architecture dependency requirements, computing power scale requirements, bandwidth requirements, storage requirements, and latency constraint requirements, and construct the task requirement matrix D. S3. For each heterogeneous computing node, calculate the heterogeneous adaptation score of that node for all subtasks to be assigned. The calculation method satisfies the formula: ； in, To match weight coefficients to the preset architecture, To match the preset bandwidth with weighting coefficients, To preset the storage matching weight coefficient, For nodes The average architecture matching degree with all unassigned subtasks, with a value ranging from [0,1]. For nodes The matching degree that satisfies the task's bandwidth requirements, with a value range of [0,1]. For nodes The matching degree that satisfies the task storage requirements, with a value range of [0,1]. For nodes Real-time power consumption per unit of computing power. For nodes Current average task queuing latency; S4, Heterogeneous Adaptation Score Construct the total cost function for the collaborative allocation of heterogeneous computing power. ; S5. Based on preset constraints, the total cost function is... The minimum value is calculated to obtain the initial collaborative allocation scheme. The constraints include: the total computing power requirement of each subtask is satisfied, the total load of each node does not exceed the maximum load threshold of the node, and the queuing delay of the allocation node for the latency-sensitive subtask does not exceed its own latency constraint. S6. Calculate the dynamic allocation correction coefficient based on the task operation parameters collected in real time during the operation cycle after the computing power allocation. ; S7, when When the value exceeds the preset correction trigger threshold, based on The current allocation scheme is partially adjusted, only the low-priority tasks carried by nodes with adaptation scores below the preset adaptation threshold are adjusted, to obtain the final heterogeneous computing power collaborative allocation scheme.

2. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S1, the feature modeling of heterogeneous computing power nodes across the network further includes: labeling cross-domain heterogeneous computing power nodes with identity tags, the tags including domain attributes, trust level, and payment level. The constraints also include: high-paying level tasks are preferentially assigned to nodes with a trust level not lower than a preset trust threshold.

3. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S2, the decomposition of the batch task set to be allocated specifically involves: constructing a directed acyclic graph based on the task dependencies, decomposing the tasks into multiple independent sub-tasks, with each sub-task corresponding to an independent computing power allocation unit.

4. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, The , , The weights are dynamically adjusted based on the overall type of the task set. When the task set is an AI inference task set, the weights are increased. The weight value is increased when the task set is a data transmission task set. The weight value of γ is increased when the task set is a big data storage task set.

5. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S4, the total cost function satisfies the formula: ； in, For nodes Normalized weights based on computing power scale For subtasks If the migration delay caused by migrating out of the initially assigned node The migration cost penalty coefficient, The fragmentation rate of the entire network's computing power after allocation. This is the fragmentation rate penalty coefficient.

6. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S5, the total cost function The minimum value is solved using a genetic algorithm. The population size is set to twice the number of heterogeneous computing power nodes, the crossover probability is 0.7, the mutation probability is 0.05, and the maximum number of iterations is 1000. The iteration is terminated when the change in the total cost function is less than 1e-3 after 50 consecutive iterations, thus obtaining the optimal initial allocation scheme.

7. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S6, the dynamic allocation correction coefficient satisfies the formula: ； in, The preset correction step size coefficient has a value range of (0,1]. This represents the average task completion latency for all assigned tasks within the current runtime cycle. This is the preset system average latency threshold.

8. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, In S7, the preset correction trigger threshold is 1.1, when When the value is greater than 1.1, an adjustment is triggered. During the adjustment process, only low-priority tasks are migrated from high-load nodes to suitable nodes with idleness higher than the preset idle threshold. The allocation position of high-priority tasks is not adjusted.

9. The heterogeneous computing power collaborative allocation method according to claim 1, characterized in that, The low-priority tasks are non-real-time tasks whose task latency constraints are greater than a preset system average latency threshold, while the high-priority tasks are real-time tasks whose task latency constraints are not greater than a preset system average latency threshold.

10. A heterogeneous computing power collaborative allocation system, characterized in that, It includes a heterogeneous computing power modeling module, a task requirement decomposition module, an allocation calculation module, a solution module, and a dynamic adjustment module. The heterogeneous computing power modeling module is connected to the task requirement decomposition module, the task requirement decomposition module is connected to the allocation calculation module, the allocation calculation module is connected to the solution module, and the solution module is connected to the dynamic adjustment module. The heterogeneous computing power modeling module is used to construct the heterogeneous computing power feature matrix; The task requirement decomposition module is used to decompose a batch of task sets and construct a task requirement matrix. The allocation calculation module is used to calculate the heterogeneous adaptation score respectively. Cooperative allocation of total cost function and dynamic allocation correction coefficient ; The solution module is used to solve the total cost function. The initial collaborative allocation scheme is obtained by solving for the minimum value. The dynamic adjustment module is used to locally adjust the current allocation scheme according to the dynamic allocation correction coefficient to obtain the final heterogeneous computing power collaborative allocation scheme.

Images