Heterogeneous GPU cluster scheduling method and system

Efficient allocation of GPU resources between heterogeneous nodes through ant colony algorithm and GPU virtualization technology, solving the problem that traditional scheduling algorithms cannot effectively utilize heterogeneous node resources, and improving the resource utilization and computing efficiency of GPU clusters.

CN118939418APending Publication Date: 2024-11-12BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410970923.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional GPU scheduling algorithms cannot effectively utilize GPU resources in heterogeneous environments, especially when the child nodes are configured differently, resulting in inadequate resource utilization and computational efficiency.

Method used

Using ant colony algorithm and GPU virtualization technology, initialize the pheromone values ​​and heuristic information values ​​of GPUs and applications, select the appropriate GPUs to allocate application resources through probability, and dynamically adjust the pheromone volatile coefficient to optimize resource allocation, and summarize the configuration of subnodes Carry out global scheduling.

Benefits of technology

It realizes efficient GPU resource scheduling between heterogeneous nodes, and improves the resource utilization and computing efficiency of the GPU cluster.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939418A_ABST
    Figure CN118939418A_ABST
Patent Text Reader

Abstract

The invention provides a heterogeneous GPU cluster scheduling method, which relates to the technical field of GPU cluster scheduling, and comprises the steps of initializing GPU and ant colony algorithm parameters, acquiring GPU scheduling indexes, initializing pheromones and heuristic information, calculating GPU allocation probability and allocating application tasks, updating pheromone values, and optimizing global scheduling. And distributing calculation data to the idle sub-nodes for calculation by utilizing a GPU virtualization technology, and finally returning a result to the GPU main node and applying a layer program. According to the method, efficient GPU resource scheduling is realized, and the resource utilization rate and the calculation efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of GPU cluster scheduling, and in particular to a heterogeneous GPU cluster scheduling method and system. Background Art

[0002] As the demand for high-performance computing increases, efficient utilization of GPU resources in clusters becomes critical. Traditional scheduling algorithms often fail to optimally utilize GPU resources in heterogeneous environments, especially when subnodes have different configurations. Therefore, a new scheduling method is proposed to improve the resource utilization and computing efficiency of GPU clusters. Summary of the invention

[0003] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a heterogeneous GPU cluster scheduling method and system, which can efficiently allocate GPU resources between heterogeneous nodes with different configurations, and the master node summarizes the configuration of the sub-nodes and performs scheduling.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A heterogeneous GPU cluster scheduling method, comprising:

[0006] Initialize the GPU and ant colony algorithm parameters, and obtain the GPU scheduling indicators;

[0007] Initialize the pheromone value of each GPU and application combination pair, and initialize the heuristic information value according to the application requirements;

[0008] Based on the pheromone value and the heuristic information value, the probability of the GPU being allocated to each application is calculated, and the application is selected according to the probability to allocate GPU resources to construct an initial solution;

[0009] The pheromone value of each GPU and application combination in the solution space of the initial solution is evaporated, and a new pheromone value is deposited according to the current resource allocation to update the state of the solution space;

[0010] Summarize the configuration of each GPU's sub-nodes, and perform global scheduling based on the results of the ant colony algorithm optimization. Use GPU virtualization technology to distribute computing data to sub-nodes with idle resources for computing, and return the results to the GPU's main node after the calculation is completed. The GPU's main node then returns the calculation results to the application layer program.

[0011] Preferably, the parameters for initializing the GPU include: GPU Max Clock rate and Memory; the parameters for initializing the ant colony algorithm include: pheromone matrix and heuristic information matrix.

[0012] Preferably, the scheduling indicators include: GPU load rate, task priority and energy efficiency.

[0013] Preferably, the process of summarizing the configurations of the subnodes of each GPU also includes:

[0014] Build a virtual graphics card matrix and integrate the GPU resources of sub-nodes.

[0015] Preferably, the update formula of the new pheromone value is: τ ij (t+1)=(1-ρ)·τ ij (t)+Δτ ij (t); where τ ij (t+1) represents the pheromone concentration on path ij at time t+1, ρ is the pheromone volatility coefficient, Δτ ij (t) is the amount of pheromone newly deposited on path ij at time t; the pheromone volatility coefficient ρ is dynamically adjusted to ensure faster response to changes under high load and stable resource allocation under low load.

[0016] Preferably, when under high load, the calculation formula of the pheromone volatility coefficient ρ is: ρ = ρ base +Δρ1, where ρ base is the basic volatility coefficient, Δρ1 is the increment under high load; when under low load, the calculation formula of pheromone volatility coefficient ρ is: ρ=ρ base -Δρ2; where Δρ2 is the increment under high load conditions.

[0017] Preferably, the calculation formula of the heuristic information value is: Among them, η ij represents the heuristic information value, d ij is the distance between node i and node j, r ij is the resource requirement of application j, u i is the real-time utilization of GPU node i, and the weight parameters ω1, ω2, and ω3 are used to balance the impact of distance, resource demand, and utilization on the heuristic information.

[0018] Preferably, the formula for allocating GPU resources by selecting applications according to probability is: P ij (t) represents the probability of transferring from node i to node j at time t, α and β are the parameters of pheromone importance and heuristic information importance respectively, and allowed is the set of nodes allowed to be selected.

[0019] A heterogeneous GPU cluster scheduling system, comprising:

[0020] The first initialization unit is used to initialize the parameters of the GPU and the ant colony algorithm and obtain the scheduling index of the GPU;

[0021] A second initialization unit is used to initialize the pheromone value of each GPU and application combination pair and initialize the heuristic information value according to the application requirements;

[0022] An initial solution construction unit, used to calculate the probability of GPU allocation to each application based on the pheromone value and the heuristic information value, and select the application according to the probability to allocate GPU resources to construct an initial solution;

[0023] An updating unit, used to evaporate the pheromone value of each combination pair of GPU and application in the solution space of the initial solution, and deposit new pheromone values ​​according to the current resource allocation situation to update the state of the solution space;

[0024] The scheduling unit is used to summarize the configuration of each GPU's sub-nodes and perform global scheduling based on the results of the ant colony algorithm optimization. It uses GPU virtualization technology to distribute computing data to sub-nodes with idle resources for computing, and returns the results to the GPU's main node after the calculation is completed. The GPU's main node then returns the calculation results to the application layer program.

[0025] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0026] The present invention provides a heterogeneous GPU cluster scheduling method and system, the method comprising: initializing the parameters of GPU and ant colony algorithm, and obtaining the scheduling index of GPU; initializing the pheromone value of each combination pair consisting of GPU and application, and initializing the heuristic information value according to the application demand; calculating the probability of GPU allocation to each application based on the pheromone value and the heuristic information value, and selecting the application according to the probability to allocate GPU resources to construct an initial solution; evaporating the pheromone value of each combination pair consisting of GPU and application in the solution space of the initial solution, and depositing new pheromone values ​​according to the current resource allocation situation to update the state of the solution space; summarizing the configuration of the sub-nodes of each GPU, and performing global scheduling according to the optimization result of the ant colony algorithm, using GPU virtualization technology to distribute the calculation data to the sub-nodes with idle resources for calculation, and returning the result to the main node of GPU after the calculation is completed, and the main node of GPU returns the calculation result to the application layer program. The present invention realizes efficient GPU resource scheduling between heterogeneous nodes, and improves the resource utilization rate and calculation efficiency of GPU cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0028] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a technical route provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] The purpose of the present invention is to provide a heterogeneous GPU cluster scheduling method and system, which can realize efficient GPU resource scheduling between heterogeneous nodes and improve the resource utilization and computing efficiency of the GPU cluster.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a heterogeneous GPU cluster scheduling method, comprising:

[0034] Step 100: Initialize the parameters of the GPU and the ant colony algorithm, and obtain the scheduling index of the GPU;

[0035] Step 200: Initializing the pheromone value of each GPU and application combination pair, and initializing the heuristic information value according to the application requirements;

[0036] Step 300: Based on the pheromone value and the heuristic information value, the probability of GPU allocation to each application is calculated, and the application is selected according to the probability to allocate GPU resources to construct an initial solution;

[0037] Step 400: evaporating the pheromone value of each GPU and application combination pair in the solution space of the initial solution, and depositing a new pheromone value according to the current resource allocation situation to update the state of the solution space;

[0038] Step 500: Summarize the configuration of the sub-nodes of each GPU, and perform global scheduling based on the results of the ant colony algorithm optimization. Use GPU virtualization technology to distribute the computing data to the sub-nodes with idle resources for computing, and return the results to the GPU master node after the calculation is completed. The GPU master node then returns the calculation results to the application layer program.

[0039] Specifically, step 100 includes initializing GPU parameters such as Max Clock rate and Memory, and ant colony algorithm parameters such as pheromone matrix and heuristic information matrix. These initialization steps provide basic data for subsequent scheduling methods, reflecting the adaptability to different GPU configurations. In step 200, the initialization of pheromone values ​​and heuristic information values ​​is performed for different GPU and application requirements, reflecting the system's support for heterogeneous environments. Step 300 is based on the pheromone value and the heuristic information value, calculating the probability of GPU allocation to each application, and selecting applications according to the probability to allocate GPU resources to build an initial solution. In this process, the system will calculate and select the most suitable GPU for resource allocation according to the characteristics and task requirements of different GPUs, further reflecting the heterogeneous characteristics. Step 400 is to evaporate the pheromone value of each combination pair of GPU and application in the solution space of the initial solution, and deposit new pheromone values ​​according to the current resource allocation situation to update the state of the solution space; the dynamic adjustment and update mechanism of pheromone supports the heterogeneity of different GPU nodes, which can adapt to different computing requirements and GPU resource states in this way. Step 500 is to summarize the configuration of the sub-nodes of each GPU, and perform global scheduling based on the results of the ant colony algorithm optimization, and use GPU virtualization technology to distribute the computing data to the sub-nodes with idle resources for computing, and return the results to the GPU master node after the calculation is completed. The GPU master node then returns the calculation results to the application layer program; at this stage, the system uses GPU virtualization technology to realize the distribution of computing data and the return of results, which enables heterogeneous GPU nodes to work together effectively and fully utilize the advantages of GPUs with different configurations.

[0040] like Figure 2 As shown, the technical route of this embodiment is as follows:

[0041] 1. System initialization:

[0042] Initialize the GPU parameters, including GPU Max Clock rate and Memory resource configuration; initialize the relevant parameters of the ant colony algorithm, such as pheromone matrix, heuristic information matrix, etc. This step provides basic configuration data for subsequent GPU resource scheduling.

[0043] 2. GPU configuration:

[0044] Obtain resource parameters such as the ID, GPU Max Clock rate, and Memory of each GPU; obtain scheduling-related indicators such as GPU load rate, task priority, and energy efficiency in real time. In the GPU configuration module, the system assigns a unique ID to each GPU and records its parameters such as GPU Max Clock rate and Memory. In addition, the system obtains scheduling-related indicators such as GPU load rate (GPU-Util), task priority, and energy efficiency (Pwr) in real time, which will be used to optimize the scheduling strategy.

[0045] Specifically, the method for obtaining the relevant indicators in this embodiment is as follows:

[0046] ① GPU load rate (GPU-Util): By monitoring the utilization of each GPU, the current GPU load is obtained in real time. This indicator is obtained by regularly querying the GPU driver or using GPU performance monitoring tools (such as NVIDIA's nvidia-smi).

[0047] ②Task priority: The task priority is specified by the application when submitting the task, indicating the importance of the task. The system determines the urgency of task scheduling based on the priority set by the application. The application uses a set of predefined priority levels, such as high, medium, and low. When submitting a task, the application passes the priority as a parameter to the scheduling system.

[0048] ③Power efficiency (Pwr): By monitoring the power consumption of each GPU, the energy efficiency of the GPU when executing a specific task is calculated. This indicator is obtained through GPU power consumption monitoring tools (such as NVIDIA's NVML API).

[0049] In addition, the indicator utilization method in this embodiment is as follows:

[0050] ① Comprehensive evaluation: The system comprehensively considers indicators such as GPU load rate, task priority, and energy efficiency to evaluate the current status and availability of each GPU node.

[0051] ②Optimize the scheduling strategy: According to the comprehensive evaluation results, adjust the scheduling strategy. For high-priority tasks, the system gives priority to GPU nodes with low load rate and high energy efficiency.

[0052] ③ Dynamic adjustment: During operation, the system monitors the changes of various indicators in real time and dynamically adjusts the scheduling strategy according to the changes to ensure optimal utilization of resources and efficient completion of tasks.

[0053] 3. Pheromone and heuristic information initialization:

[0054] Initialize pheromone values ​​for each GPU-application pair; initialize heuristic information values ​​based on the resource requirements of the application. The heuristic information is dynamically adjusted to reflect the task characteristics (such as computation-intensive, memory-intensive, etc.) and the real-time status of the GPU node.

[0055] 4. Construct solution space:

[0056] In the solution space construction module, based on the initialized pheromone and heuristic information, the system calculates the probability of each GPU being assigned to each application. New dimensions such as task priority and energy efficiency are introduced, so that each task can find the GPU node that best suits its characteristics.

[0057] Through the probabilistic selection mechanism of the ant colony algorithm, the system selects appropriate applications according to the calculated probability to allocate GPU resources and build an initial solution.

[0058] 5. Pheromone update:

[0059] The pheromone update module evaporates the pheromone of each GPU-application pair in the solution space, indicating that the influence of the original pheromone gradually weakens over time.

[0060] According to the current resource allocation, new pheromones are deposited and the state of the solution space is updated. According to the actual effect of task execution, the pheromones are adjusted. The allocation scheme with short execution time and high resource utilization will get more pheromone deposition, thus guiding subsequent ants to prefer these paths.

[0061] Dynamically adjust the pheromone evaporation rate to ensure that the system can respond to changes more quickly when the load is high and maintain stable resource allocation when the load is low.

[0062] The pheromone update formula is as follows:

[0063] τ ij (t+1)=(1-ρ)·τ ij (t)+Δτ ij (t)

[0064] Among them, τ ij (t+1) represents the pheromone concentration on path ij at time t+1, ρ is the pheromone volatility coefficient, Δτ ij (t) is the amount of pheromone newly deposited on path ij at time t. The pheromone volatility coefficient ρ is adjusted dynamically to ensure that the system can respond to changes more quickly when the load is high and maintain stable resource allocation when the load is low.

[0065] The dynamic adjustment of the pheromone volatility coefficient is as follows:

[0066] Adjustments for high load conditions:

[0067] When the system detects a high load (such as GPU utilization exceeding 80%), the system increases the pheromone volatility coefficient ρ to speed up the evaporation rate of the pheromone in order to adapt to new task allocation requirements more quickly.

[0068] Adjustment formula: ρ high =ρ base +Δρ, where ρ base is the basic volatility coefficient and Δρ is the increment under high load conditions.

[0069] Adjustment for low load conditions:

[0070] When the system detects low load (such as GPU utilization is less than 20%), the system reduces the pheromone volatility coefficient ρ to slow down the evaporation rate of pheromones and maintain the effectiveness of the current task allocation strategy.

[0071] Adjustment formula: ρ low =ρ base -Δρ, where ρ base is the basic volatility coefficient and Δρ is the reduction under low load conditions.

[0072] Heuristic Information Formula

[0073] The heuristic information formula is as follows:

[0074]

[0075] Among them, η ij represents heuristic information, d ij is the distance between node i and node j, r ij is the resource requirement of application j, u i is the real-time utilization of GPU node i. The weight parameters ω1, ω2, ω3 are used to balance the impact of distance, resource demand and utilization on the heuristic information. The heuristic information is initialized according to the resource demand of the application and dynamically adjusted to reflect the task characteristics and the real-time status of the GPU node.

[0076] Probability selection formula:

[0077] The probability selection formula is as follows:

[0078]

[0079] Among them, P ij (t) represents the probability of transferring from node i to node j at time t, α and β are the parameters of pheromone importance and heuristic information importance respectively, and allowed is the set of nodes allowed to be selected.

[0080] The new formula combining task priority and energy efficiency is as follows:

[0081]

[0082] Among them, Priority(j) represents the priority of task j, and EnergyEfficiency(j) represents the energy efficiency of node j. This formula introduces the dimensions of task priority and energy efficiency, so that each task can find the GPU node that best suits its characteristics.

[0083] The new formula combining task priority and energy efficiency in this embodiment is used to calculate the GPU allocation probability, and task priority and energy efficiency are included in the calculation, so as to select the most suitable GPU node.

[0084] During the pheromone update process, the dynamically adjusted pheromone volatility coefficient also takes into account task priority and energy efficiency to ensure that resource allocation can be optimized under different load conditions. Through these steps, the heterogeneous GPU cluster scheduling method achieves efficient resource allocation and optimization, and the new formula combining task priority and energy efficiency further improves the accuracy and efficiency of scheduling.

[0085] 6. Scheduling execution:

[0086] The scheduling execution module is the core part of the entire system. The master node summarizes the configuration of each child node and performs global scheduling based on the results of the improved ant colony algorithm optimization. Combining historical data and current load conditions, a machine learning model is used to predict the future load of each GPU node, helping ants to select nodes more intelligently when building paths.

[0087] GPU virtualization technology is used to distribute computing data to sub-nodes with idle resources for computing. After the sub-nodes complete the computing, the computing results are returned to the main node. The main node returns the computing results of the sub-nodes to the application layer program to complete the computing tasks and output the results. Through GPU virtualization technology, the system can dynamically allocate and manage GPU resources, significantly improving resource utilization and system performance.

[0088] In addition, this embodiment also discloses a heterogeneous GPU cluster scheduling system based on an improved ant colony algorithm and GPU virtualization technology, including:

[0089] System initialization module, used to initialize GPU parameters, including GPU Max Clock rate and Memory resource configuration;

[0090] GPU configuration module, used to configure parameters such as ID, GPU Max Clock rate, Memory, GPU load rate, task priority, and energy efficiency of each GPU;

[0091] The pheromone and heuristic information initialization module is used to initialize the pheromone value for each GPU and initialize the heuristic information value according to the application requirements;

[0092] The solution space construction module is used to calculate the probability of GPU allocation to each application based on pheromone and heuristic information, and construct the initial solution according to the probability selection;

[0093] The pheromone update module is used to evaporate the pheromone of each GPU-application pair in the solution space and deposit new pheromone according to the allocation;

[0094] The scheduling execution module is used to summarize the configuration of each child node and perform global scheduling based on the results of the ant colony algorithm optimization. It uses GPU virtualization technology to distribute the computing data to child nodes with idle resources for computing, and returns the results to the main node after the calculation is completed. The main node then returns the calculation results to the application layer program.

[0095] Corresponding to the above method, this embodiment also provides another heterogeneous GPU cluster scheduling system, including:

[0096] The first initialization unit is used to initialize the parameters of the GPU and the ant colony algorithm and obtain the scheduling index of the GPU;

[0097] A second initialization unit is used to initialize the pheromone value of each GPU and application combination pair and initialize the heuristic information value according to the application requirements;

[0098] An initial solution construction unit, used to calculate the probability of GPU allocation to each application based on the pheromone value and the heuristic information value, and select the application according to the probability to allocate GPU resources to construct an initial solution;

[0099] An updating unit, used to evaporate the pheromone value of each combination pair of GPU and application in the solution space of the initial solution, and deposit new pheromone values ​​according to the current resource allocation situation to update the state of the solution space;

[0100] The scheduling unit is used to summarize the configuration of each GPU's sub-nodes and perform global scheduling based on the results of the ant colony algorithm optimization. It uses GPU virtualization technology to distribute computing data to sub-nodes with idle resources for computing, and returns the results to the GPU's main node after the calculation is completed. The GPU's main node then returns the calculation results to the application layer program.

[0101] The beneficial effects of the present invention are as follows:

[0102] The present invention can efficiently allocate GPU resources for applications between heterogeneous nodes with different configurations, and the master node summarizes the configuration of the child nodes and performs global scheduling. The system uses GPU virtualization technology to enable GPU resources to be flexibly allocated, and optimizes the scheduling strategy through the ant colony algorithm to improve cluster computing efficiency. The method includes the steps of system initialization, GPU configuration, initialization of pheromone and heuristic information, construction of solution space, pheromone update and scheduling execution, which realizes efficient GPU resource scheduling between heterogeneous nodes and improves the resource utilization and computing efficiency of the GPU cluster.

[0103] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0104] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A heterogeneous GPU cluster scheduling method, characterized in that: include: Initialize the GPU and ant colony algorithm parameters, and obtain the GPU scheduling indicators; Initialize the pheromone value of each GPU and application combination pair, and initialize the heuristic information value according to the application requirements; Based on the pheromone value and the heuristic information value, the probability of the GPU being allocated to each application is calculated, and the application is selected according to the probability to allocate GPU resources to construct an initial solution; The pheromone value of each GPU and application combination in the solution space of the initial solution is evaporated, and a new pheromone value is deposited according to the current resource allocation to update the state of the solution space; Summarize the configuration of each GPU's sub-nodes, and perform global scheduling based on the results of the ant colony algorithm optimization. Use GPU virtualization technology to distribute computing data to sub-nodes with idle resources for computing, and return the results to the GPU's main node after the calculation is completed. The GPU's main node then returns the calculation results to the application layer program.

2. The heterogeneous GPU cluster scheduling method according to claim 1, characterized in that: The parameters for initializing the GPU include: GPU Max Clock rate and Memory; the parameters for initializing the ant colony algorithm include: pheromone matrix and heuristic information matrix.

3. The heterogeneous GPU cluster scheduling method according to claim 1, characterized in that: The scheduling indicators include: GPU load rate, task priority and energy efficiency.

4. The heterogeneous GPU cluster scheduling method according to claim 1, characterized in that: The process of summarizing the configuration of each GPU's subnodes also includes: Build a virtual graphics card matrix and integrate the GPU resources of sub-nodes.

5. The heterogeneous GPU cluster scheduling method according to claim 1, characterized in that: The update formula of the new pheromone value is: τ ij (t+1)=(1-ρ)·τ ij (t)+Δτ ij (t); where τ ij (t+1) represents the pheromone concentration on path ij at time t+1, ρ is the pheromone volatility coefficient, Δτ ij (t) is the amount of pheromone newly deposited on path ij at time t; the pheromone volatility coefficient ρ is dynamically adjusted to ensure faster response to changes under high load and stable resource allocation under low load.

6. The heterogeneous GPU cluster scheduling method according to claim 5, characterized in that: When under high load, the calculation formula of pheromone volatility coefficient ρ is: ρ = ρ base +Δρ1, where ρ base is the basic volatility coefficient, Δρ1 is the increment under high load; when under low load, the calculation formula of pheromone volatility coefficient ρ is: ρ=ρ base -Δρ2; where Δρ2 is the increment under high load conditions.

7. The heterogeneous GPU cluster scheduling method according to claim 5, characterized in that: The calculation formula of the heuristic information value is: Among them, η ij represents the heuristic information value, d ij is the distance between node i and node j, r ij is the resource requirement of application j, u i is the real-time utilization of GPU node i, and the weight parameters ω1, ω2, and ω3 are used to balance the impact of distance, resource demand, and utilization on the heuristic information.

8. The heterogeneous GPU cluster scheduling method according to claim 7, characterized in that: The formula for allocating GPU resources by selecting applications according to probability is: P ij (t) represents the probability of transferring from node i to node j at time t, α and β are the parameters of pheromone importance and heuristic information importance respectively, and allowed is the set of nodes allowed to be selected.

9. A heterogeneous GPU cluster scheduling system, characterized in that: include: The first initialization unit is used to initialize the parameters of the GPU and the ant colony algorithm and obtain the scheduling index of the GPU; A second initialization unit is used to initialize the pheromone value of each GPU and application combination pair and initialize the heuristic information value according to the application requirements; An initial solution construction unit, used to calculate the probability of GPU allocation to each application based on the pheromone value and the heuristic information value, and select the application according to the probability to allocate GPU resources to construct an initial solution; An updating unit, used to evaporate the pheromone value of each combination pair of GPU and application in the solution space of the initial solution, and deposit new pheromone values ​​according to the current resource allocation situation to update the state of the solution space; The scheduling unit is used to summarize the configuration of each GPU's sub-nodes and perform global scheduling based on the results of the ant colony algorithm optimization. It uses GPU virtualization technology to distribute computing data to sub-nodes with idle resources for computing, and returns the results to the GPU's main node after the calculation is completed. The GPU's main node then returns the calculation results to the application layer program.

Citation Information

Cited By

  • Self-adaptive adjustment method and system for production parameters of steel blanks

    CN119989573A

  • Method and system for scheduling heterogeneous GPU (Graphics Processing Unit) by cross-cluster management software

    CN121542052A