Intelligent agent computing power scheduling method for holographic resource perception

Abstract

The invention discloses a holographic resource sensing agent computing power scheduling method, which comprises the following steps of S1, dividing computing power resources into a plurality of resource domains, and deploying an agent in each resource domain for managing the computing power resources in the corresponding resource domain; s2, obtaining calculation tasks submitted by a user, extracting key features of the calculation tasks, and classifying the calculation tasks according to the key features to obtain calculation task types; s3, states of computing power resources in the domain are monitored through the agents, scheduling strategies in all the agents are generated according to the types of the computing tasks and the states of the computing power resources, the computing power resources are distributed to the computing tasks according to the scheduling strategies, and the computing tasks submitted by the user are completed; and S4, monitoring task execution states including a resource utilization rate and a task success rate, sending the task execution states to the intelligent agent, and driving iterative optimization of the scheduling strategy library. According to the method, the resource state is monitored through the intelligent agent and matched with the task requirement, the dynamic scheduling strategy is generated and executed, and the task scheduling flexibility and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to an intelligent agent computing power scheduling method based on holographic resource perception. Background Technology

[0002] With the rapid development of technology, the demand for computing resources is increasing daily, especially in fields such as artificial intelligence, big data analysis, and high-performance computing. Traditional computing resource scheduling methods are no longer sufficient to meet the diverse and complex needs of modern computing tasks. Existing power grid dispatch centers typically employ centralized scheduling for computing resource management. This approach, when faced with large-scale, heterogeneous computing resources, cannot provide refined management of different types of computing resources, resulting in low resource utilization. Furthermore, existing scheduling methods struggle to dynamically schedule based on key characteristics of computing tasks, such as computational complexity, real-time requirements, and resource dependencies, thus failing to effectively match task demands with resource capabilities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for intelligent agent computing power scheduling based on holographic resource perception.

[0004] The holographic resource perception-based intelligent agent computing power scheduling method provided by this invention includes the following steps:

[0005] S1: Divide computing resources into multiple resource domains, and deploy an intelligent agent in each resource domain to manage the computing resources within the corresponding resource domain;

[0006] S2: Obtain the computing tasks submitted by the user, extract the key features of the computing tasks, classify the computing tasks according to the key features, and obtain the computing task type.

[0007] S3: Monitor the status of computing resources within the domain through intelligent agents, generate scheduling strategies for each intelligent agent based on the type of computing task and the status of computing resources, and allocate computing resources to computing tasks according to the scheduling strategies to complete the computing tasks submitted by the user.

[0008] S4: Monitor task execution status, including resource utilization and task success rate, and send the data to the agent to drive iterative optimization of the scheduling strategy library.

[0009] Preferably, in step S1, specifically,

[0010] S11: The computing power network is divided into multiple resource domains based on the hardware type of the computing power resources, including CPU resource domain, GPU resource domain, FPGA resource domain and memory-intensive resource domain;

[0011] S12: Deploy an agent within each resource domain to monitor and manage the status of computing resources within the domain, including computing power, storage capacity, and network bandwidth.

[0012] Preferably, step S2 specifically involves:

[0013] S21: Receive a computing task request submitted by the user, and extract the key features of the task, including: computing complexity features, real-time requirement features, and resource dependency features.

[0014] S22: Classify computing tasks based on key features to obtain computing task types, which include high-performance computing tasks, real-time computing tasks, general-purpose computing tasks, and special-purpose acceleration tasks.

[0015] S23: Generate a corresponding resource requirement description based on the task type. The resource requirement description includes computing resource requirements, storage resource requirements, and network resource requirements.

[0016] Preferably, in step S3, specifically,

[0017] S31: Monitor the status indicators of computing resources in each resource domain through intelligent agents in each resource domain, including the real-time utilization rate of computing resources, the available capacity of storage resources, and the current bandwidth and latency of network resources.

[0018] S32: Match the monitored resource status with the resource requirement description corresponding to the computing task type, and calculate the fit score for each resource domain.

[0019] S33: Based on the fitness score, each resource domain agent generates a scheduling strategy, executes resource allocation according to the scheduling strategy, and completes the computation task submitted by the user.

[0020] Preferably, step S32 specifically includes:

[0021] S321: Obtain the computing resource requirements, storage resource requirements, and network resource requirements from the resource requirement description corresponding to the computing task type;

[0022] S322: Obtain real-time utilization of computing resources, available capacity of storage resources, and current bandwidth and latency of network resources for each resource domain as monitored by each agent;

[0023] S323: The computing resource matching degree is obtained by calculating the matching degree between the remaining computing power of each resource domain and the computing task requirements;

[0024] S324: The storage resource matching degree is obtained by calculating the matching degree between the available capacity of storage resources in each resource domain and the calculated storage resource requirements;

[0025] S325: The network resource matching degree is obtained by calculating the matching degree between the network resources of each resource domain and the network resource requirements of the computing tasks;

[0026] S326: The resource domain fit score is calculated based on the matching degree of computing resources, storage resources and network resources and the preset weights.

[0027] Preferably, step S33 specifically includes:

[0028] S331: Sort the adaptation scores of each resource domain from high to low, and select candidate resource domains whose adaptation scores are higher than the preset threshold.

[0029] S332: Select the resource domain with the highest fit score from the candidate resource domains and determine it as the target resource domain to be allocated;

[0030] S333: Send a resource allocation request to the target resource domain. The request includes the specific configuration and quantity of computing resources, storage resources and network resources required by the task.

[0031] S334: After receiving a resource allocation request, the target resource domain allocates the corresponding resources according to its own resource status and task requirements and locks them for task execution;

[0032] S335: If the target resource domain successfully allocates resources and starts the task, the agent sends a message to the user that the task execution has started.

[0033] S336: If the target resource domain cannot satisfy the resource allocation request, the agent will reselect the next resource domain with the highest fitness score from the candidate resource domains to try to allocate resources, until the resource is successfully allocated or all candidate resource domains have been tried.

[0034] Preferably, step S4 specifically includes:

[0035] S41: After the task is completed, obtain the deviation rate between the actual consumption of computing resources and the estimated value.

[0036] A status indicator indicating whether the task was completed within the preset time;

[0037] S42: Aggregate data according to a fixed time window to generate a resource scheduling performance report;

[0038] S43: When the task success rate is lower than the first preset threshold or the resource utilization deviation is greater than the second preset threshold, the parameter tuning of the scheduling strategy library is triggered.

[0039] This invention discloses a holographic resource perception-based intelligent agent computing power scheduling method, which has the following beneficial effects:

[0040] By dividing resource domains according to hardware type and deploying dedicated intelligent agents, refined management and control of heterogeneous resources can be achieved; based on the classification of key task features, multi-dimensional resource requirement descriptions are generated, and the intelligent agents monitor resource status and match it with task requirements, realizing the generation and execution of dynamic scheduling strategies, thus improving the flexibility and accuracy of task scheduling.

[0041] By monitoring task execution status, a resource scheduling performance report is generated, and parameter tuning of the scheduling strategy library is triggered when the task success rate falls below a threshold or the resource utilization deviation exceeds a threshold. This mechanism ensures that the scheduling strategy can be continuously optimized based on actual operating conditions, thereby improving task success rate and resource utilization.

[0042] By calculating resource matching degree, storage resource matching degree, and network resource matching degree, and combining them with preset weights, this invention can accurately evaluate the suitability score of each resource domain. This suitability calculation method can effectively match task requirements with resource capabilities, ensuring that tasks are assigned to the most suitable resource domain, thereby improving task execution efficiency. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart of the intelligent agent computing power scheduling method for holographic resource perception provided by the present invention;

[0045] Figure 2 The flowchart of step S3 in the intelligent agent computing power scheduling method for holographic resource perception provided by the present invention;

[0046] Figure 3 The flowchart of step S32 in the intelligent agent computing power scheduling method for holographic resource perception provided by the present invention;

[0047] Figure 4 The flowchart of step S33 in the intelligent agent computing power scheduling method for holographic resource perception provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0049] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0050] refer to Figure 1 The holographic resource perception-based intelligent agent computing power scheduling method provided by this invention includes the following steps:

[0051] S1: Divide computing resources into multiple resource domains, and deploy an intelligent agent in each resource domain to manage the computing resources within the corresponding resource domain;

[0052] S2: Obtain the computing tasks submitted by the user, extract the key features of the computing tasks, classify the computing tasks according to the key features, and obtain the computing task type.

[0053] S3: Monitor the status of computing resources within the domain through intelligent agents, generate scheduling strategies for each intelligent agent based on the type of computing task and the status of computing resources, and allocate computing resources to computing tasks according to the scheduling strategies to complete the computing tasks submitted by the user.

[0054] S4: Monitor the task execution status, including resource utilization and task success rate, and send the data to the agent to drive iterative optimization of the scheduling strategy library.

[0055] The holographic resource-aware intelligent agent computing power scheduling method provided by this invention can achieve refined management and control of heterogeneous resources by dividing resource domains according to hardware type and deploying intelligent agents; it generates multi-dimensional resource demand descriptions based on task key feature state classification, and matches resource status with task requirements through intelligent agents to realize the generation and execution of dynamic scheduling strategies, thereby improving the flexibility and accuracy of task scheduling; by monitoring the task execution status, it drives the iterative optimization of the scheduling strategy library, thereby improving the task success rate and resource utilization.

[0056] In a preferred embodiment, step S1 specifically involves:

[0057] S11: The computing power network is divided into multiple resource domains based on the hardware type of the computing power resources, including CPU resource domain, GPU resource domain, FPGA resource domain and memory-intensive resource domain;

[0058] S12: Deploy an agent within each resource domain to monitor and manage the status of computing resources within the domain, including computing power, storage capacity, and network bandwidth.

[0059] In the preferred embodiment, step S2 specifically involves:

[0060] S21: Receive a computing task request submitted by the user, and extract the key features of the task, including: computing complexity features, real-time requirement features, and resource dependency features.

[0061] S22: Classify computing tasks based on key features to obtain computing task types, which include high-performance computing tasks, real-time computing tasks, general-purpose computing tasks, and special-purpose acceleration tasks.

[0062] The classification of computational tasks based on key features includes:

[0063] When the computational complexity exceeds the first threshold, it is classified as computationally intensive.

[0064] When the computational complexity characteristic is below the first threshold, it is classified as non-computationally intensive.

[0065] When the task completion deadline is lower than the second threshold, it is classified as real-time sensitive.

[0066] When the task completion deadline exceeds the second threshold, it is classified as a non-real-time sensitive type.

[0067] Based on the extracted resource dependency features, the task's dependency on specific hardware is identified, including: CPU instruction set dependency, GPU acceleration library dependency, FPGA logic unit dependency, and memory bandwidth dependency.

[0068] Based on the above classification results, they are mapped to high-performance computing tasks, real-time computing tasks, general computing tasks, and special acceleration tasks through predefined rules;

[0069] For example, high-performance computing tasks: computationally intensive and non-real-time sensitive tasks, dedicated accelerated tasks: tasks that rely on GPU / FPGA hardware acceleration capabilities, etc.

[0070] S23: Generate a corresponding resource requirement description based on the task type. The resource requirement description includes computing resource requirements, storage resource requirements, and network resource requirements.

[0071] For example, the resource requirement descriptions generated for high-performance computing tasks include:

[0072] Computing resource requirements: Multi-core CPU topology requirements and minimum clock speed requirements;

[0073] Storage resource requirements: shared memory capacity and memory access bandwidth threshold;

[0074] Network resource requirements: RDMA network support identifiers and maximum latency thresholds;

[0075] The resource requirements description generated for real-time computing tasks includes:

[0076] Computing resource requirements: minimum single-core CPU clock speed and cache capacity requirements;

[0077] Storage resource requirements: In-memory database access interface requirements;

[0078] Network resource requirements: deterministic bandwidth guarantee and jitter limit;

[0079] The resource requirements description generated for dedicated accelerated tasks includes:

[0080] Computing resource requirements: Accelerator type (GPU / FPGA / TPU) and video memory / HBM capacity;

[0081] Storage resource requirements: Data exchange rate between device memory and host memory;

[0082] Network resource requirements: GPU-Direct or InfiniBand support identifier;

[0083] The resource requirements description generated for general-purpose computing tasks includes:

[0084] Computing resource requirements: Virtualization container specifications;

[0085] Storage resource requirements: Baseline block storage IOPS;

[0086] Network resource requirements: minimum bandwidth guarantee.

[0087] refer to Figure 2 In a preferred embodiment, step S3 specifically involves:

[0088] S31: Monitor the status indicators of computing resources in each resource domain through intelligent agents in each resource domain, including the real-time utilization rate of computing resources, the available capacity of storage resources, and the current bandwidth and latency of network resources.

[0089] S32: Match the monitored resource status with the resource requirement description corresponding to the computing task type, and calculate the fit score for each resource domain.

[0090] S33: Based on the fitness score, each resource domain agent generates a scheduling strategy, executes resource allocation according to the scheduling strategy, and completes the computation task submitted by the user.

[0091] refer to Figure 3In a preferred embodiment, step S32 specifically includes:

[0092] S321: Obtain the computing resource requirements, storage resource requirements, and network resource requirements from the resource requirement description corresponding to the computing task type;

[0093] S322: Obtain real-time utilization of computing resources, available capacity of storage resources, and current bandwidth and latency of network resources for each resource domain as monitored by each agent;

[0094] S323: The computing resource matching degree is obtained by calculating the matching degree between the remaining computing power of each resource domain and the computing task requirements;

[0095] The calculation formula is as follows:

[0096]

[0097] C max To maximize the computing power of the resource domain, C used For the computing power already used in the resource domain, C task The task requires computational resources.

[0098] S324: The storage resource matching degree is obtained by calculating the matching degree between the available capacity of storage resources in each resource domain and the calculated storage resource requirements;

[0099] The calculation formula is as follows:

[0100]

[0101] Among them, S available S represents the available storage capacity of the resource domain. task To meet the resource requirements for task storage.

[0102] S325: The network resource matching degree is obtained by calculating the matching degree between the network resources of each resource domain and the network resource requirements of the computing tasks;

[0103] The calculation formula is as follows:

[0104]

[0105] B threshold B is the bandwidth threshold. current L represents the current bandwidth of the resource domain. threshold L is the delay threshold. current Current delay for the resource domain;

[0106] S326: The resource domain fit score is calculated based on the matching degree of computing resources, storage resources and network resources and the preset weights.

[0107] The calculation formula is as follows:

[0108] S=αMc+βMs+γMn

[0109] Where α, β, and γ are the weight coefficients of each resource matching degree, and α+β+γ=1.

[0110] This embodiment calculates resource matching degree, storage resource matching degree, and network resource matching degree, and combines them with preset weights. The present invention can accurately evaluate the adaptability score of each resource domain. This adaptability calculation method can effectively match task requirements with resource capabilities, ensuring that tasks are assigned to the most suitable resource domain, thereby improving task execution efficiency.

[0111] refer to Figure 4 In a preferred embodiment, step S33 specifically includes:

[0112] S331: Sort the adaptation scores of each resource domain from high to low, and select candidate resource domains whose adaptation scores are higher than the preset threshold.

[0113] S332: Select the resource domain with the highest fit score from the candidate resource domains and determine it as the target resource domain to be allocated;

[0114] S333: Send a resource allocation request to the target resource domain. The request includes the specific configuration and quantity of computing resources, storage resources and network resources required by the task.

[0115] S334: After receiving a resource allocation request, the target resource domain allocates the corresponding resources according to its own resource status and task requirements and locks them for task execution;

[0116] S335: If the target resource domain successfully allocates resources and starts the task, the agent sends a message to the user that the task execution has started.

[0117] S336: If the target resource domain cannot satisfy the resource allocation request, the agent will reselect the next resource domain with the highest fitness score from the candidate resource domains to try to allocate resources, until the resource is successfully allocated or all candidate resource domains have been tried.

[0118] If all candidate resource domains are unable to meet the resource allocation request, the agent decides whether to request cross-domain resource coordination and rescheduling based on the urgency of the task and resource requirements.

[0119] In a preferred embodiment, step S4 specifically includes:

[0120] S41: After the task is completed, obtain the deviation rate between the actual consumption of computing resources and the estimated value.

[0121] A status indicator indicating whether the task was completed within the preset time;

[0122] S42: Aggregate data according to a fixed time window to generate a resource scheduling performance report;

[0123] S43: When the task success rate is lower than the first preset threshold or the resource utilization deviation is greater than the second preset threshold, the parameter optimization of the scheduling strategy library is triggered, such as adjusting the weight value coefficient of resource matching degree and other parameters.

Claims

1. A method for scheduling computing power of intelligent agents based on holographic resource perception, characterized in that, Includes the following steps: S1: Divide computing resources into multiple resource domains, and deploy an intelligent agent in each resource domain to manage the computing resources within the corresponding resource domain; S2: Obtain the computing tasks submitted by the user, extract the key features of the computing tasks, classify the computing tasks according to the key features, and obtain the computing task type. S3: Monitor the status of computing resources within the domain through intelligent agents, generate scheduling strategies for each intelligent agent based on the type of computing task and the status of computing resources, and allocate computing resources to computing tasks according to the scheduling strategies to complete the computing tasks submitted by the user. S4: Monitor task execution status, including resource utilization and task success rate, and send the data to the agent to drive iterative optimization of the scheduling strategy library.

2. The intelligent agent computing power scheduling method for holographic resource perception according to claim 1, characterized in that, In step S1, specifically, S11: The computing power network is divided into multiple resource domains based on the hardware type of the computing power resources, including CPU resource domain, GPU resource domain, FPGA resource domain and memory-intensive resource domain; S12: Deploy an agent within each resource domain to monitor and manage the status of computing resources within the domain, including computing power, storage capacity, and network bandwidth.

3. The intelligent agent computing power scheduling method for holographic resource perception according to claim 1, characterized in that, Step S2, specifically, S21: Receive a computing task request submitted by the user, and extract the key features of the task, including: computing complexity features, real-time requirement features, and resource dependency features. S22: Classify computing tasks based on key features to obtain computing task types, which include high-performance computing tasks, real-time computing tasks, general-purpose computing tasks, and special-purpose acceleration tasks. S23: Generate a corresponding resource requirement description based on the task type. The resource requirement description includes computing resource requirements, storage resource requirements, and network resource requirements.

4. The intelligent agent computing power scheduling method for holographic resource perception according to claim 1, characterized in that, In step S3, specifically, S31: Monitor the status indicators of computing resources in each resource domain through intelligent agents in each resource domain, including the real-time utilization rate of computing resources, the available capacity of storage resources, and the current bandwidth and latency of network resources. S32: Match the monitored resource status with the resource requirement description corresponding to the computing task type, and calculate the fit score for each resource domain. S33: Based on the fitness score, each resource domain agent generates a scheduling strategy, executes resource allocation according to the scheduling strategy, and completes the computation task submitted by the user.

5. The intelligent agent computing power scheduling method for holographic resource perception according to claim 4, characterized in that, Step S32 specifically includes: S321: Obtain the computing resource requirements, storage resource requirements, and network resource requirements from the resource requirement description corresponding to the computing task type; S322: Obtain real-time utilization of computing resources, available capacity of storage resources, and current bandwidth and latency of network resources for each resource domain as monitored by each agent; S323: The computing resource matching degree is obtained by calculating the matching degree between the remaining computing power of each resource domain and the computing task requirements; S324: The storage resource matching degree is obtained by calculating the matching degree between the available capacity of storage resources in each resource domain and the calculated storage resource requirements; S325: The network resource matching degree is obtained by calculating the matching degree between the network resources of each resource domain and the network resource requirements of the computing tasks; S326: The resource domain fit score is calculated based on the matching degree of computing resources, storage resources and network resources and the preset weights.

6. The intelligent agent computing power scheduling method for holographic resource perception according to claim 5, characterized in that, Step S33 specifically includes: S331: Sort the adaptation scores of each resource domain from high to low, and select candidate resource domains whose adaptation scores are higher than the preset threshold. S332: Select the resource domain with the highest fit score from the candidate resource domains and determine it as the target resource domain to be allocated; S333: Send a resource allocation request to the target resource domain. The request includes the specific configuration and quantity of computing resources, storage resources and network resources required by the task. S334: After receiving a resource allocation request, the target resource domain allocates the corresponding resources according to its own resource status and task requirements and locks them for task execution; S335: If the target resource domain successfully allocates resources and starts the task, the agent sends a message to the user that the task execution has started. S336: If the target resource domain cannot satisfy the resource allocation request, the agent reselects the next resource domain with the highest fitness score from the candidate resource domains to try to allocate resources, until the resource is successfully allocated or all candidate resource domains have been tried.

7. The intelligent agent computing power scheduling method for holographic resource perception according to claim 1, characterized in that, Step S4 specifically includes: S41: After the task is completed, obtain the deviation rate of the actual consumption of computing resources compared with the estimated value and the status marker of whether the task was completed within the preset time. S42: Aggregate data according to a fixed time window to generate a resource scheduling performance report; S43: When the task success rate is lower than the first preset threshold or the resource utilization deviation is greater than the second preset threshold, the parameter tuning of the scheduling strategy library is triggered.

Images