An intelligent computing power scheduling method and system based on heterogeneous resource pooling

Abstract

This invention relates to an intelligent computing power scheduling method and system based on heterogeneous resource pooling. The invention integrates metrics such as computing power, GPU memory, storage performance, and network latency, and calculates a load factor through dynamic weights (adjusted according to task type) to quantify the overall load status of nodes. It optimizes scheduling weights in real time using reinforcement learning, and constructs a node comprehensive scoring model by combining node load factors, task adaptability (the degree of matching between task requirements and node resources), and stability scores. Based on the node comprehensive score, a multi-objective constraint optimization model is introduced to balance resource requirements and performance objectives, supporting dynamic adjustment of task priorities and cross-cluster resource aggregation, achieving multi-objective scheduling optimization of "load balancing—task adaptability—stability assurance." This invention provides a standardized solution for the efficient utilization of heterogeneous computing power resources and is adaptable to diverse scenarios such as AI training and real-time inference.

Description

Technical Field

[0001] This invention belongs to the field of computing power scheduling technology, and in particular relates to an intelligent computing power scheduling method and system based on heterogeneous resource pooling. Background Technology

[0002] With the rapid evolution of scenarios such as artificial intelligence training, autonomous driving inference, and scientific computing, the demand for computing power is experiencing explosive growth, and the composition of computing resources is becoming increasingly diversified, with heterogeneous hardware becoming the norm. However, traditional computing power management models face three core contradictions in addressing this trend:

[0003] First, the fragmented deployment of heterogeneous hardware leads to resource fragmentation. Devices from different manufacturers are difficult to coordinate due to differences in driver interfaces, forming "hardware islands" with generally low resource utilization.

[0004] Secondly, the resource requirements of tasks are becoming more refined and dynamic (such as 1% computing power granularity and 1MB video memory granularity requirements), but the existing system only supports whole card or simple proportional partitioning, which cannot match the diverse needs, resulting in both idle resources and contention for resources.

[0005] Third, the scheduling decision relies on static rules and does not combine the real-time status of nodes with the task type for dynamic optimization, which leads to load imbalance. Some nodes experience a sharp drop in performance due to overload, while others remain idle for a long time. Summary of the Invention

[0006] To address the aforementioned issues, this invention aims to provide an intelligent computing power scheduling method and system based on heterogeneous resource pooling, forming a closed-loop management mechanism of "perception-pooling-scheduling-execution," providing a standardized solution for the efficient utilization of heterogeneous computing power resources, and adapting to diverse scenarios such as AI training and real-time inference.

[0007] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides an intelligent computing power scheduling method based on heterogeneous resource pooling, comprising:

[0008] The system collects core metrics from each computing node in real time and performs normalization processing. These core metrics include at least computing power usage, video memory usage, disk I / O rate, and inter-node communication latency.

[0009] The normalization results are weighted and summed to obtain the load factor used to quantify the overall load status of the nodes.

[0010] Construct and solve the following multi-objective constrained optimization model to obtain the task allocation scheme:

[0011] , , ,

[0012] In the formula, x i Let x be a binary variable. i =1 indicates that the task is assigned to node i, x i =0 indicates that no task is assigned to node i; S i R is the overall score for node i; j,i R represents the remaining amount of the j-th type of resource at node i; j,t Let n be the total requirement of task t for the j-th type of resource; n is the total number of nodes;

[0013] Based on the task allocation scheme, tasks are distributed to nodes to achieve computing power scheduling.

[0014] As a further improvement to one embodiment of the present invention, the calculation of the load factor includes,

[0015] ,

[0016] In the formula, Let be the load factor of node i. These are the normalized current GPU / NPU computing power usage, current video memory usage, disk I / O rate, and inter-node communication performance of node i, respectively. All are weighting coefficients.

[0017] As a further improvement to one embodiment of the present invention, the calculation of the comprehensive score of a node includes,

[0018] ,

[0019] In the formula, All are dynamic weights; The fit between task t and node i; Score the stability of node i.

[0020] As a further improvement to one embodiment of the present invention, the calculation of the task-node fit includes,

[0021] ,

[0022] In the formula, Let i be the total capacity of the j-th type of resource. Let be the total capacity of the j-th type of resources in the system.

[0023] As a further improvement to one embodiment of the present invention, the stability calculation of the node includes,

[0024] ,

[0025] In the formula, Let the standard deviation of the task completion delay for node i be . Let be the average completion delay of the task at node i. Let $\frac{i}{i}$ be the standard deviation of the task completion time fluctuation for node $i$. Let be the average completion time of the task at node i.

[0026] As a further improvement to one embodiment of the present invention, a reinforcement learning mechanism is used to adjust the dynamic weights, wherein the reward function is: , For throughput, Waiting time for the task This is the throughput weighting coefficient. This represents the penalty coefficient for waiting time.

[0027] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides an intelligent computing power scheduling system based on heterogeneous resource pooling, comprising:

[0028] The resource control module is used to collect the core indicators of each computing node in the system in real time and perform normalization processing; the normalization processing results are weighted and summed to obtain the load factor used to quantify the comprehensive load status of the computing nodes.

[0029] The intelligent scheduling module is used to construct and solve a multi-objective constrained optimization model to obtain a task allocation scheme;

[0030] The task distribution module is used to distribute tasks to nodes based on the task allocation scheme, thereby realizing the scheduling of computing power.

[0031] As a further improvement to one embodiment of the present invention, the system further includes,

[0032] The computing power pooling module is used to form atomic capability units through standardized computing power modeling and abstraction. It supports fine-grained scheduling of resources in two dimensions: absolute value partitioning and proportional partitioning, and builds a unified computing power resource pool.

[0033] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps in the intelligent computing power scheduling method based on heterogeneous resource pooling as described above.

[0034] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps in the intelligent computing power scheduling method based on heterogeneous resource pooling as described above.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) Innovative dual-dimensional resource allocation mode (absolute value allocation and proportional allocation) supports fine allocation of 1% computing power and 1MB video memory granularity, significantly improving resource utilization;

[0037] (2) Establish a multi-dimensional scheduling model that integrates load factor, task adaptability and node stability score, and dynamically optimizes scheduling weight through reinforcement learning to achieve load balancing and efficient task matching.

[0038] (3) Form a closed-loop management mechanism of "perception-pooling-scheduling-execution" to provide a standardized solution for the efficient utilization of heterogeneous computing resources and adapt to diverse scenarios such as AI training and real-time inference. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the system framework described in this invention. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0041] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0042] In Embodiment 1 of the present invention, an intelligent computing power scheduling system based on heterogeneous resource pooling is provided, such as... Figure 1 As shown, it includes:

[0043] Computing power pooling module: Enables unified management and dynamic adaptation of computing power devices from multiple vendors. By abstracting atomic capability units through standardized computing power modeling, it supports fine-grained resource allocation in two dimensions: absolute value allocation (e.g., 1% computing power, 1MB video memory) and proportional allocation (e.g., 1 / 4 card), and builds a unified computing power resource pool.

[0044] Resource control module: Real-time collection of core indicators such as computing power, video memory, storage performance, network latency, and hardware status; calculates load factors through normalization processing; and quantifies the overall load status of nodes by combining dynamic threshold management and alarm mechanisms.

[0045] Intelligent scheduling module: Based on node load factor, task adaptability and stability score, a multi-dimensional scoring model is constructed. Reinforcement learning is combined to dynamically optimize scheduling weights. A multi-objective constraint optimization model is introduced to balance resource demand and performance goals and dynamically select the optimal resource node.

[0046] Task distribution module: Implements task distribution and deployment through containerization technology (such as Kubernetes), supports multi-protocol communication and full lifecycle status tracking, integrates dynamic adjustment, fault tolerance mechanism and security permission control to ensure efficient task execution.

[0047] 1. Computing power pooling module

[0048] 1.1 Unified Management of Heterogeneous Resources

[0049] To address the management barriers and scheduling challenges of heterogeneous GPUs (NVIDIA, Huawei Ascend, Hygon DCU, etc.), a three-tiered collaborative system of "cluster-device-plugin" is constructed to achieve unified access across clusters for computing devices from multiple vendors. Dynamic plug-in distribution overcomes driver barriers and supports automatic discovery and state synchronization.

[0050] (1) Cluster and device management: Supports one-click import of K8S clusters (including number of nodes and authentication information), automatically scans node GPU devices through the unified Device Plugin, and also supports manually specifying cluster, node and plugin version to add devices, realizing unified awareness of cross-cluster and cross-vendor resources (single cluster management response ≤ 5 seconds, adapts to ≥ 3 types of heterogeneous GPUs).

[0051] (2) Unified adaptation of heterogeneous GPUs: A global centralized plug-in controller is built by integrating components such as HAMI and vDTK to dynamically distribute plug-ins from multiple vendors and break through the barriers of hardware driver differences; through software virtualization technology, it supports fine-grained partitioning of 1% computing power / 1MB video memory and realizes dynamic isolation and monitoring of GPU resources within containers.

[0052] (3) Standardized capability output: The computing power, video memory and other capabilities of heterogeneous GPUs are abstracted into standardized interfaces, shielding the differences in underlying hardware and providing a unified resource view for upper-layer scheduling.

[0053] 1.2 Standardized Computing Power Modeling System

[0054] To address the fragmentation problem in heterogeneous resource management, an integrated solution of "unified modeling—resource abstraction—capability encapsulation—model configuration" is constructed to solve the hardware collaboration challenge.

[0055] (1) By using attribute modeling (defining core attributes such as GPU computing power and video memory) + topology modeling (characterizing resource connection relationships), a metadata description framework for heterogeneous resources is formed to achieve collaborative perception of "computing, storage and network".

[0056] (2) Abstract complex hardware capabilities (such as GPU computing and network forwarding) into atomic capability units (such as 1% computing power and 1MB video memory), encapsulate them into RESTful interfaces, and complete the standardized conversion of "hardware capabilities → service interfaces".

[0057] (3) With the help of the visual configurator, real hardware models (such as A100, Ascend 910B) are mapped to logical models (such as “computing power GPU-300T”), and resource pools can be configured by “atomic capability + logical model” combination (such as “10 computing power GPUs + 5 storage nodes”) to meet the personalized needs of businesses.

[0058] This system provides a technical foundation for resource pooling that enables "unified capabilities and flexible access," supporting multi-strategy resource allocation in the intelligent scheduling layer.

[0059] 1.3 Refined Resource Allocation

[0060] It supports two dimensions of vGPU resource partitioning modes to meet the computing power requirements of different scenarios and realize flexible allocation and dynamic management of heterogeneous GPU resources.

[0061] (1) Absolute value segmentation

[0062] Based on the task's precise requirements for computing power and GPU memory, specific values ​​are directly specified for partitioning:

[0063]

[0064] Where: vC is the computing power after partitioning, vM is the video memory after partitioning; α is the computing power ratio specified for the task (1%≤α≤100%, supports 1% granular adjustment); γ is the video memory requirement specified for the task (unit: MB, supports 1MB granular adjustment).

[0065] (2) Proportional segmentation

[0066] The GPU is partitioned based on its overall physical proportion (e.g., 1 / 4 card, 25% card), with computing power and video memory allocated in equal proportions:

[0067]

[0068] Where: vC represents the computing power after partitioning, vM represents the memory after partitioning; C total M represents the total computing power of physical GPUs. total δ represents the total physical GPU memory; δ represents the percentage of the card specified for the task (e.g., 1 / 4 card, 25% card), and the percentage is usually determined by the characteristics of the physical card.

[0069] The two splitting modes mentioned above can be preset through business templates and selected as needed when submitting tasks, or they can be configured directly when submitting tasks. The vGPU resources generated after splitting are automatically associated with task IDs and can be automatically / manually released after the task is completed or in abnormal situations, ensuring efficient reuse of resources.

[0070] 2. Resource Control Module

[0071] 2.1 Collection of core indicators

[0072] The system collects the following core hardware resource metrics in real time:

[0073] (1) Computing power (C) i The current GPU / NPU computing power usage (unit: TFLOPs) directly affects the computational efficiency of a task and is a core metric for scheduling. If the node's computing power is insufficient, the task execution time will increase significantly.

[0074] (2) Video Memory (M) i ): Current video memory usage (unit: GB), which limits model size and batch size. Insufficient video memory may cause the task to fail due to OOM (Out of Memory).

[0075] (3) Storage performance (S) i Disk I / O rate (unit: IOPS) affects data read and write efficiency, especially in large-scale data training or distributed tasks, where insufficient storage performance may lead to bottlenecks in computing and I / O.

[0076] (4) Network communication performance indicators (D) i Network communication performance (RCS) is used to characterize the communication performance between nodes and can be calculated based on parameters such as communication latency, bandwidth, or packet loss rate. In distributed or real-time inference tasks, network communication performance directly affects the coordination and execution efficiency between tasks, and nodes whose network communication performance meets the task requirements should be selected first.

[0077] 2.2 Normalization of core indicators

[0078] The purpose of normalization is to eliminate the dimensional differences between different resource indicators and ensure the load factor λ. i The calculations are fair and comparable. For example, the difference in units of computing power and video memory: computing power (TFLOPs) and video memory (GB) cannot be directly compared, but after normalization, they can be unified to the [0,1] range; the difference in units of storage performance and network latency: storage performance (IOPS) and network latency (ms) have significantly different numerical ranges, and normalization can prevent high numerical indicators from dominating the load factor results.

[0079] For each resource metric, Min-Max Normalization is performed using the following formula:

[0080]

[0081] Where: X i The current indicator value (C) i M i S i D i ); X min X max These represent the minimum and maximum values ​​of the indicator (e.g., the minimum computing power is 0 TFLOPs, and the maximum is 400 TFLOPs). The normalized indicator range is [0,1], representing the resource utilization ratio: 0 indicates that the resources are completely idle; 1 indicates that the resources have reached their limit (e.g., the video memory is fully occupied).

[0082] 2.3 Definition and Calculation of Loading Factor

[0083] The load factor is used to quantify the comprehensive load status of computing nodes on multi-dimensional resources such as computing power, video memory, storage and network. Through dynamic weight adjustment and normalization, it provides a quantitative basis for the resource occupancy of task scheduling, thereby optimizing resource utilization and avoiding overload risks.

[0084] Load factor λ i The normalized resource indicators are combined, and the priorities of different resources are dynamically adjusted using weighting coefficients ω1, ω2, ω3, and ω4.

[0085]

[0086] Where: weighting coefficient ; These are normalized resource indicators.

[0087] The design of weighting coefficients ω1, ω2, ω3, and ω4 needs to be combined with the task type and system objectives.

[0088] (1) For computationally intensive tasks (such as deep learning training): increase ω1 and decrease other weights;

[0089] (2) For memory-intensive tasks (such as large-scale model inference): increase ω2 and decrease other weights;

[0090] (3) Storage performance-sensitive tasks (such as big data processing): Increase ω3 and decrease other weights;

[0091] (4) Low-latency tasks (such as real-time inference): Increase ω4 and decrease other weights.

[0092] 2.4 Resource Threshold Management and Alarm Mechanism

[0093] (1) Dynamic threshold: Set resource occupancy thresholds (e.g., λ) based on historical load distribution. i >0.8 triggers an alarm);

[0094] (2) Tiered alarms: Different levels of alarms (such as yellow / red warnings) are triggered according to the degree of exceeding the limit, and resource expansion or task migration strategies are linked.

[0095] (3) Automatic repair: When the node load continues to exceed the limit, trigger elastic resource allocation (such as Kubernetes HPA automatic scaling).

[0096] 3. Intelligent scheduling module

[0097] 3.1 Node Scoring and Task Adaptation Mechanism

[0098] The intelligent scheduling module selects the optimal node based on a multi-dimensional scoring formula, which is:

[0099]

[0100] Where: λ i The load factor is provided by the resource control module (see Equation 1); A i,t The fitness degree between task t and node i (see Equation 6); S stability,i The stability score for node i is given (see Equation 7); α, β, and γ are dynamic weights that are adaptively adjusted according to the task type.

[0101] The dynamic weight adjustment strategy is as follows:

[0102] ① Resource-priority scenarios (such as sudden high load): Increase α, decrease β, γ;

[0103] ② Task-sensitive scenarios (such as critical tasks): Increase β and prioritize matching nodes with high compatibility;

[0104] ③ For scenarios prioritizing stability (such as long-cycle tasks): Increase γ and select nodes with stable historical performance.

[0105] 3.2 Task Fit Calculation

[0106] Task adaptability A i,t The formula reflecting the matching degree between node resource distribution and task requirements is:

[0107]

[0108] Where: R j,i R represents the current value of the j-th type of resource at node i (e.g., GPU memory); j,tLet R represent the resource requirements of task t for the j-th type. j,total,i R represents the total capacity of the j-th type of resource at node i; j,total Let be the total capacity of the j-th type of resources in the system (used for normalization). The higher the fit (approaching 1), the better the resource allocation of node i matches the needs of task t.

[0109] 3.3 Node Stability Score Calculation

[0110] Stability rating S stability,i The formula for quantifying the historical performance of nodes is:

[0111]

[0112] Where: σ latency,i μ is the standard deviation of the task completion delay for node i. latency,i σ is the average completion delay of the task at node i. completion,i Let μ be the standard deviation of the task completion time fluctuation at node i. completion,i This represents the average task completion time for node i. A higher stability score (approaching 1) indicates more stable node performance.

[0113] 3.4 Dynamic Weight Optimization and Multi-Objective Scheduling

[0114] 3.4.1. Employing a reinforcement learning mechanism to dynamically adjust weights:

[0115] (1) State space: node load distribution, task queue length, and historical scheduling effect;

[0116] (2) Action space: Adjust the weight combination of α, β, and γ;

[0117] (3) Reward function: For throughput, Waiting time for the task This is the throughput weighting coefficient. Waiting time penalty coefficient:

[0118]

[0119] 3.4.2. Introduce a multi-objective constrained optimization model to maximize the sum of node scores. To achieve the goal while satisfying resource constraints and decision constraints This achieves a balance between resource feasibility and performance optimization. The multi-objective constrained optimization model is as follows:

[0120]

[0121]

[0122] Where: x i S is a binary variable (1 indicates that a task is assigned to node i, and 0 indicates that no task is assigned); i Rate the nodes (see Equation 5); R j,i R represents the remaining amount of the j-th type of resource at node i; j,t Let t be the total requirement of task t for the j-th type of resource; These correspond to the computing power, video memory, storage, and network communication performance indicators in section 2.1, respectively.

[0123] This invention optimizes scheduling weights (α, β, γ) in real time through reinforcement learning, combined with the node load factor λ. i Task adaptability A i,t (The degree of matching between task requirements and node resources) and stability score S stability,i A node comprehensive scoring model is constructed; a multi-objective constraint optimization model is introduced to balance resource demand and performance objectives, support dynamic adjustment of task priorities and cross-cluster resource aggregation, and achieve multi-objective scheduling optimization of "load balancing - task adaptation - stability guarantee".

[0124] 4. Task Issuance Module

[0125] 4.1 Task Distribution Strategy and Protocol

[0126] The task distribution module achieves efficient task scheduling through flexible distribution strategies and communication protocols. Centralized scheduling is decided uniformly by a central scheduler (such as the Kubernetes API Server), which is suitable for scenarios with clear resource requirements and significant differences in task priorities. It allocates tasks through preset scheduling strategies (such as priority ranking and resource matching). Distributed scheduling, on the other hand, uses a consistent hashing algorithm to achieve efficient mapping between tasks and nodes. It is particularly suitable for edge computing or large-scale cluster scenarios. By dynamically adjusting the hash ring weight, it balances node load and reduces cross-node communication overhead.

[0127] In terms of communication protocols, gRPC is a high-performance RPC framework based on HTTP / 2, supporting streaming transmission and bidirectional communication, and is suitable for tasks with high real-time requirements (such as AI inference); MQTT, as a lightweight message queue protocol, is designed for edge computing, and achieves efficient task delivery under low bandwidth through a topic subscription mechanism, while also supporting threshold triggering (such as automatically switching protocols to reduce latency when the node load exceeds 90%).

[0128] 4.2 Task Execution and Status Tracking

[0129] During task execution, a heartbeat mechanism and anomaly detection are used to achieve full lifecycle monitoring. The task periodically reports its status (such as runtime and resource usage). If the heartbeat interval exceeds a preset threshold (such as no report for 15 seconds), the node anomaly detection process is triggered. Anomaly detection, combined with time series analysis, identifies task anomalies (such as a sudden increase in GPU usage exceeding 80% or memory leaks) and triggers automatic repair strategies.

[0130] Status tracking records task lifecycle events (submission, execution, completion) through event logs and integrates with visualization dashboards (such as Grafana / Prometheus) to display task execution status and resource consumption trends in real time. Furthermore, the system supports dynamic threshold adjustment: when the task type changes (e.g., switching from batch processing to stream processing), the monitoring metric thresholds are automatically adjusted (e.g., adjusting the memory usage limit from 90% to 70%) to avoid false alarms or missed alarms.

[0131] 4.3 Dynamic Adjustment and Fault Tolerance Mechanism

[0132] The task distribution module supports dynamic resource adjustment and fault-tolerant recovery to ensure system robustness. Dynamic adjustment is based on the actual resource requirements of the task (such as computing power, memory, and storage) to achieve elastic scaling up and down, while introducing resource utilization thresholds and cold start optimization.

[0133] Fault tolerance mechanisms include task retry and node migration (when a node fails or the load exceeds a threshold, the task is automatically migrated to a replacement node, and the availability of the target node's resources is verified through health checks before migration).

[0134] 4.4 Security and Access Control

[0135] The task distribution module ensures system security through multi-layered security mechanisms. Access control is based on RBAC (Role-Based Access Control) to restrict task submission permissions. Different roles (such as administrators, developers, and ordinary users) have differentiated operation permissions. Identity authentication is implemented through OAuth2.0 / JWT. If the number of consecutive authentication failures exceeds a threshold (e.g., 3 times), account lockout and alarms are triggered. Regarding data security, task data is transmitted using TLS 1.3 encryption, and sensitive data (such as user credentials) is processed using regular expression replacement or field anonymization. The system also supports access frequency threshold control (e.g., a maximum of 100 tasks per minute) to prevent malicious attacks or resource abuse. Furthermore, critical operations (such as task deletion and permission changes) require secondary verification (such as SMS verification codes or biometrics) to ensure the legitimacy of the operation.

[0136] 4.5 Anomaly Handling and Optimization

[0137] 4.5.1 Load Migration: When the node load λ iWhen the score exceeds the threshold, some tasks are migrated to high-scoring nodes;

[0138] 4.5.2 Task Queue: Tasks that do not meet the requirements are put into the FIFO queue and dynamically scheduled according to priority. The order of tasks in the queue can be manually adjusted (such as moving up, moving down, or placing on top).

[0139] In Embodiment 2 of the present invention, an intelligent computing power scheduling method based on heterogeneous resource pooling is provided, including:

[0140] The system collects core metrics from each computing node in real time and performs normalization processing. These core metrics include at least computing power usage, video memory usage, disk I / O rate, and inter-node communication latency.

[0141] The normalization results are weighted and summed to obtain the load factor used to quantify the overall load status of the nodes.

[0142] Construct and solve a multi-objective constrained optimization model to obtain a task allocation scheme;

[0143] Based on the task allocation scheme, tasks are distributed to nodes to achieve computing power scheduling.

[0144] Those skilled in the art will understand that, for the sake of convenience and brevity, the steps and procedures of the above-mentioned intelligent computing power scheduling method can be referred to the specific working process of the aforementioned intelligent computing power scheduling system, and will not be repeated here.

[0145] In a third embodiment of the present invention, an electronic device is provided, including a memory and a processor. The memory stores a computer program that can run on the processor. When the program is executed on the processor, the steps in the intelligent computing power scheduling method based on heterogeneous resource pooling as described above are implemented.

[0146] In Embodiment 4 of the present invention, a storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the steps in the intelligent computing power scheduling method based on heterogeneous resource pooling as described above.

[0147] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the modules described above can be referred to the corresponding process in the aforementioned method implementation, and will not be repeated here.

[0148] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0149] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.

[0150] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer system (which may be a personal computer, server, or network system, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A smart computing power scheduling method based on heterogeneous resource pooling, characterized in that: include, The system collects core metrics from each computing node in real time and performs normalization processing. These core metrics include at least computing power usage, video memory usage, disk I / O rate, and inter-node communication latency. The normalization results are weighted and summed to obtain the load factor used to quantify the overall load status of the nodes. Construct and solve the following multi-objective constrained optimization model to obtain the task allocation scheme: ， ， ， In the formula, x i Let x be a binary variable. i =1 indicates that the task is assigned to node i, x i =0 indicates that no task is assigned to node i; S i R is the overall score for node i; j,i R represents the remaining amount of the j-th type of resource at node i; j,t Let n be the total requirement of task t for the j-th type of resource; n is the total number of nodes; Based on the task allocation scheme, tasks are distributed to nodes to achieve computing power scheduling.

2. The method according to claim 1, characterized in that: The calculation of the load factor includes, ， In the formula, Let be the load factor of node i. These are the normalized current GPU / NPU computing power usage, current video memory usage, disk I / O rate, and inter-node communication performance of node i, respectively. All are weighting coefficients.

3. The method according to claim 2, characterized in that: The calculation of the node's overall score includes, ， In the formula, All are dynamic weights; The fit between task t and node i; Score the stability of node i.

4. The method according to claim 3, characterized in that: The calculation of task-node fit includes, ， In the formula, Let i be the total capacity of the j-th type of resource. Let be the total capacity of the j-th type of resources in the system.

5. The method according to claim 3, characterized in that: The stability calculation of a node includes, ， In the formula, Let the standard deviation of the task completion delay for node i be . Let be the average completion delay of the task at node i. Let $\frac{i}{i}$ be the standard deviation of the task completion time fluctuation for node $i$. Let be the average completion time of the task at node i.

6. The method according to claim 3, characterized in that: A reinforcement learning mechanism is used to adjust the dynamic weights, where the reward function is... , For throughput, Waiting time for the task This is the throughput weighting coefficient. This represents the penalty coefficient for waiting time.

7. A system applying the method as described in any one of claims 1 to 6, characterized in that: include, The resource control module is used to collect the core indicators of each computing node in the system in real time and perform normalization processing; the normalization processing results are weighted and summed to obtain the load factor used to quantify the comprehensive load status of the computing nodes. The intelligent scheduling module is used to construct and solve a multi-objective constrained optimization model to obtain a task allocation scheme; The task distribution module is used to distribute tasks to nodes based on the task allocation scheme, thereby realizing the scheduling of computing power.

8. The system according to claim 7, characterized in that: It also includes, The computing power pooling module is used to form atomic capability units through standardized computing power modeling and abstraction. It supports fine-grained scheduling of resources in two dimensions: absolute value partitioning and proportional partitioning, and builds a unified computing power resource pool.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps of the method as described in any one of claims 1-6.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images