Computing power resource scheduling method and system for AI multi-service data center

Abstract

The invention relates to the technical field of artificial intelligence data centers, and discloses a computing power resource scheduling method and system for an AI multi-service data center, and the method comprises the following steps: S1, obtaining multi-service task request data, processing the task request data based on a time sequence analysis model, and generating resource demand prediction data; s2, according to the resource demand prediction data, heterogeneous computing resource data is converted into virtual resource pool data, and priority queue data is generated based on task priority data; s3, processing the virtual resource pool data and the task priority data based on a reinforcement learning algorithm to generate resource allocation strategy data, and processing high-frequency access data through a prefetching strategy; and S4, generating edge-cloud hierarchical scheduling data according to the task delay demand data and the network topology data. According to the method, heterogeneous resources are dynamically allocated through time sequence analysis and a reinforcement learning algorithm, the resource utilization rate is increased to 85% from 60%, and the vacancy rate is reduced to 5%.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence data centers, and particularly to a computing power resource scheduling method and system for an AI multi-service data center. Background Art

[0002] With the rapid development of artificial intelligence technology, AI multi-service data centers need to support diverse tasks, such as AI model training, inference, scientific computing, and commercial rendering. In the prior art, traditional schedulers (such as Slurm) achieve task scheduling through static resource allocation and simple priority strategies, allocate computing resources (such as CPUs, GPUs) to tasks based on predefined rules, and manage the task execution order through a priority queue. These methods have achieved resource allocation and task scheduling to a certain extent and are applicable to single-machine or small-scale cluster environments.

[0003] However, traditional scheduling methods have significant deficiencies in high-concurrency and heterogeneous resource scenarios, especially in terms of resource utilization. Static resource allocation is difficult to adapt to the dynamic changes in task requirements, resulting in resource contention or idle during high-concurrency periods (such as the peak of night inference tasks). The resource utilization rate is usually only about 60%, and the idle rate is as high as more than 20%. For example, when processing diverse AI tasks, fixed allocation rules cannot be adjusted in real time according to task priorities and resource status, causing high-priority tasks (such as scientific research computing) to be preempted by low-priority tasks, with a low success rate of task allocation and limited overall system efficiency. Summary of the Invention

[0004] To make up for the above deficiencies, the present invention provides a computing power resource scheduling method and system for an AI multi-service data center, aiming to improve the problem that traditional static resource allocation is difficult to adapt to the dynamic changes in task requirements, resulting in resource contention or idle during high-concurrency periods, with the resource utilization rate usually only about 60% and the idle rate as high as more than 20%.

[0005] In a first aspect, the present invention provides the following technical solution. A computing power resource scheduling method for an AI multi-service data center includes the following steps: S1. Obtain multi-service task request data, process the task request data based on a time series analysis model, and generate resource demand prediction data; S2. According to the resource demand prediction data, convert heterogeneous computing resource data into virtual resource pool data, generate priority queue data based on task priority data, and process the priority queue data through containerization technology for resource isolation and dynamic allocation; S3. Process the virtual resource pool data and task priority data based on a reinforcement learning algorithm to generate resource allocation policy data. Meanwhile, process the high-frequency access data through a prefetch policy to generate optimized data locality data, so as to balance resource utilization, energy consumption, and task latency; S4. Generate edge-cloud hierarchical scheduling data according to the task latency requirement data and network topology data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined networking to generate optimized transmission data.

[0006] Through the above technical solutions, in S1, the LSTM model is used to analyze task request data and predict resource requirements; in S2, resource pooling and dynamic allocation are achieved through KVM virtualization and Kubernetes; in S3, the DDPG reinforcement learning algorithm is adopted to optimize resource allocation, and combined with the Redis prefetch policy to reduce data access latency; in S4, tasks are scheduled based on the SDN and edge-cloud architecture to optimize cross-node transmission. This method significantly improves resource utilization and task efficiency when dealing with high-concurrency tasks and is applicable to scenarios such as AI training and inference.

[0007] Preferably, it further includes: S5. Process node status data based on a dual-active data center architecture to generate failover data, and combine checkpoint technology to process long-cycle task data to generate task continuity data; S6. Process task allocation data based on real-time load data to generate load balancing data to avoid local resource overload or idleness; S7. Generate green scheduling data according to task type data and power usage efficiency data, allocate non-real-time task data to low power usage efficiency node data, and process processor frequency data through dynamic voltage and frequency scaling technology, and combine defragmentation algorithms to process fragmented resource data to generate energy efficiency optimization data.

[0008] Preferably, the time series analysis model in step S1 includes an autoregressive integrated moving average model or a long short-term memory network model, and the resource demand prediction data includes the predicted usage data of the central processing unit, graphics processing unit, and storage resources within a future period of time.

[0009] Preferably, the reinforcement learning algorithm in step S3 adopts the deep deterministic policy gradient algorithm to process the virtual resource pool data and task priority data to generate resource allocation policy data. The prefetch policy processes high-frequency access data through a distributed cache system to generate optimized data locality data and reduce cross-node data transmission latency.

[0010] Second aspect, the present invention provides the following technical solution: A computing power resource scheduling system for an AI multi-service data center, the system comprising: A resource management module, configured to obtain multi-service task request data, process the task request data based on a time series analysis model, and generate resource demand prediction data; A resource allocation module, configured to convert heterogeneous computing resource data into virtual resource pool data according to the resource demand prediction data, generate priority queue data based on task priority data, and process the priority queue data through containerization technology for resource isolation and dynamic allocation; A scheduling optimization module, configured to process the virtual resource pool data and task priority data based on a reinforcement learning algorithm to generate resource allocation policy data, and at the same time process high-frequency access data through a prefetch policy to generate optimized data locality data; A cross-region scheduling module, configured to generate edge-cloud hierarchical scheduling data according to task latency requirement data and network topology data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined network to generate optimized transmission data.

[0011] Preferably, it further comprises: A fault tolerance management module, configured to process node status data based on a dual-active data center architecture to generate failover data, and process long-period task data in combination with checkpoint technology to generate task continuity data; A load balancing module, configured to process task allocation data based on real-time load data to generate load balancing data; An energy efficiency optimization module, configured to generate green scheduling data according to task type data and power usage efficiency data, and process processor frequency data and fragmented resource data through dynamic voltage and frequency adjustment technology and defragmentation algorithm to generate energy efficiency optimization data.

[0012] Preferably, the reinforcement learning algorithm in the scheduling optimization module includes a deep deterministic policy gradient algorithm, which is used to process the virtual resource pool data and task priority data to generate resource allocation policy data, and the prefetch policy processes high-frequency access data through a distributed cache system to generate optimized data locality data.

[0013] Preferably, it further comprises: A monitoring and management module, configured to obtain resource utilization data, storage input / output performance data, and energy consumption data, process the data based on a data center infrastructure management system, and generate visual resource situation data.

[0014] In a third aspect, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above-mentioned computing power resource scheduling method for an AI multi-service data center is implemented.

[0015] In a fourth aspect, the invention provides the following technical solution: a readable storage medium with a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned computing power resource scheduling method for an AI multi-service data center is implemented.

[0016] The invention has the following beneficial effects: 1. In the present invention, heterogeneous resources (such as CPUs and GPUs) are dynamically allocated through time series analysis and reinforcement learning algorithms. Compared with traditional static scheduling, the resource utilization rate is increased from 60% to 85%, and the idle rate is reduced to 5%. In a scenario of 1,000 concurrent AI tasks, the priority queue and containerization technology avoid resource contention, and the allocation success rate reaches 95%, supporting the efficient execution of AI training and rendering tasks.

[0017] 2. In the present invention, edge-cloud hierarchical scheduling and SDN are adopted to optimize cross-node transmission, reducing the transmission delay by 50% and achieving a storage I / O throughput of 10 GB / s. The completion time of a distributed AI training task (10 TB dataset) is shortened from 48 hours to 32 hours, with a 40% efficiency improvement, making it suitable for scenarios such as autonomous driving and large-scale model training.

[0018] 3. In the present invention, short-term task scheduling is optimized through reinforcement learning and prefetching strategies, reducing the cold start delay from 500 ms to 200 ms, a 40% reduction. In a scenario of 1,000 inference tasks, the cache hit rate reaches 90%, and the access delay is reduced to 10 ms, improving the response speed of video stream processing and real-time inference.

[0019] 4. In the present invention, the failover time < 1 second and the recovery time < 10 seconds are achieved through a dual-active data center and checkpoint technology. When testing 100,000 tasks, the switching success rate is 100%, and the system availability reaches 99.99%, meeting the high-reliability requirements of long-cycle tasks such as gene sequencing. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flowchart of a method for a computing power resource scheduling method for an AI multi-service data center proposed by the present invention; Figure 2 is a system design architecture diagram of a computing power resource scheduling system for an AI multi-service data center proposed by the present invention; Figure 3This is the configuration architecture diagram of the computing power center device for a computing power resource scheduling system for an AI multi-service data center proposed by the present invention. Detailed implementation manners

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment 1 Refer to Figures 1-3 , in the first embodiment of the present invention, the present invention provides a computing power resource scheduling method for an AI multi-service data center, including the following steps: S1. Obtain multi-service task request data, process the task request data based on a time series analysis model, and generate resource demand prediction data; S2. According to the resource demand prediction data, convert heterogeneous computing resource data into virtual resource pool data, generate priority queue data based on task priority data, and process the priority queue data through containerization technology for resource isolation and dynamic allocation; S3. Process the virtual resource pool data and task priority data based on a reinforcement learning algorithm, generate resource allocation policy data, and at the same time process high-frequency access data through a prefetch policy to generate optimized data locality data to balance resource utilization, energy consumption, and task latency; S4. Generate edge-cloud hierarchical scheduling data according to task latency requirement data and network topology data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined networking to generate optimized transmission data.

[0023] Specifically, in a specific implementation, step S1 receives task request data from a multi-service client through a server. The task request data includes task types (such as AI model training, inference, scientific computing), task priorities, resource requirements (such as the number of CPU cores, GPU video memory, storage capacity), and latency requirements. The time series analysis model deployed on the server uses a long short-term memory network (LSTM). Its input is the resource usage data (including CPU utilization, GPU occupancy, storage I / O throughput) and task request frequency in the past 24 hours, and the output is the resource demand prediction data for the next 1 hour, with a prediction accuracy of over 90%. For example, in an AI training scenario, the LSTM model analyzes historical data and predicts that the required GPU resources during peak periods (such as a surge in night-time inference tasks) are 2000GB of video memory, generating resource demand prediction data. Step S2 uses virtualization technology (such as KVM) to convert heterogeneous computing resources (including Intel Xeon CPUs, NVIDIA A100 GPUs, NVMe SSD storage) into a virtual resource pool with a unified interface. The resource pool data includes the total available resources and the allocation status. The task priority data is divided into three levels according to business requirements: high (scientific computing), medium (commercial rendering), and low (test tasks). Resource isolation is achieved through Kubernetes namespaces (Namespaces) to prevent high-priority tasks from being preempted by low-priority tasks. For example, a high-priority AI training task is assigned an independent Namespace and exclusively occupies 80% of the GPU resources. Step S3 uses the deep deterministic policy gradient (DDPG) reinforcement learning algorithm, which takes the virtual resource pool data (total resources, allocation status) and task priority data as inputs and outputs resource allocation policy data (such as allocating 4 GPU cores to high-priority tasks). The DDPG algorithm optimizes the allocation through a reward function (resource utilization 0.5 + energy consumption 0.3 + latency * 0.2), converges after 1000 training iterations, and the resource utilization is increased by 35%. The prefetch policy is based on the Redis cache system, pre-storing frequently accessed data (such as AI model parameters) in memory, and reducing the access latency from 50ms to 10ms. Step S4 generates edge-cloud hierarchical scheduling data based on the task latency requirements (low-latency tasks < 10ms, large-scale training tasks < 100ms) and network topology data (bandwidth and latency between nodes). For example, an autonomous driving data processing task is assigned to an edge node (latency 5ms), and a Transformer model training task is assigned to the cloud center (bandwidth 100Gbps). The software-defined network (SDN) uses an InfiniBand switch, configures the RoCE protocol, and the cross-node transmission rate reaches 200Gbps, with the latency reduced by 50%. In practical applications, this method processes 1000 concurrent tasks in an AI data center, with a resource utilization rate of 85% and the task completion time shortened by 40%.

[0024] It also includes: S5. Process the node status data based on the active-active data center architecture to generate failover data, and combine the checkpoint technology to process the long-cycle task data to generate task continuity data; S6. Process the task allocation data based on the real-time load data to generate load balancing data, avoiding local resource overload or idleness; S7. Generate green scheduling data according to the task type data and the power usage efficiency data, allocate the non-real-time task data to the nodes with low power usage efficiency, and process the processor frequency data through the dynamic voltage and frequency scaling technology, and combine the disk defragmentation algorithm to process the fragmented resource data to generate energy efficiency optimization data.

[0025] Specifically, step S5 realizes high availability through the active-active data center architecture. The server collects the node status data in real time (including CPU / GPU load, network connection status, storage availability), and detects faults through the heartbeat mechanism. For example, after a certain node loses power, the primary data center (located in the east) switches to the standby data center (located in the west) within 1 second, and the Zookeeper coordination service ensures consistency during the switching process. The checkpoint technology saves the intermediate state of the long-cycle tasks (such as gene sequencing tasks) to the Ceph distributed storage every 5 minutes, occupying only 5MB / s of storage bandwidth, ensuring that the task can be restored to the breakpoint after the task continuity data is generated, and the recovery time < 10 seconds. Step S6 collects the real-time load data (updated once per second) through the Prometheus monitoring system, including the CPU utilization rate (0-100%) of each node, the GPU video memory occupancy (0-80GB), and the storage I / O throughput (0-10GB / s). The load balancing algorithm is based on the principle of the minimum number of connections, and allocates new tasks to the node with the lowest load. For example, when the load of node A is 90% and the load of node B is 30%, the new task is allocated to node B, and the local overload rate is reduced by 60% after generating the load balancing data. Step S7 generates green scheduling data according to the task type data (real-time tasks such as video stream processing, non-real-time tasks such as batch rendering) and the PUE data (PUE of the eastern node = 1.5, PUE of the western node = 1.1), and preferentially allocates non-real-time tasks to the western nodes with low PUE, and the scheduling decision time < 100ms. The dynamic voltage and frequency scaling (DVFS) technology dynamically adjusts the CPU frequency (1.2GHz - 3.8GHz) according to the task load. For example, it is reduced to 1.5GHz under low load, saving 20% of energy consumption. The disk defragmentation algorithm scans the fragmented resources (such as 1GB of GPU video memory, 2 CPU cores), and integrates them into allocable units through the greedy algorithm, supporting the rapid execution of small tasks (such as inference tasks), and reducing the resource fragmentation rate by 70%. In a one-month test, this method processes 100,000 tasks, the failover success rate is 100%, the energy consumption drops by 25%, and the PUE is stable below 1.2.

[0026] The time series analysis model in step S1 includes an autoregressive integrated moving average model or a long short-term memory network model, and the resource demand prediction data includes the predicted usage data of the central processing unit, graphics processing unit, and storage resources within a future period of time.

[0027] Specifically, the time series analysis model in step S1 is specifically implemented as an LSTM network. The network structure includes 3 hidden layers (each layer has 128 neurons). The input features are historical resource usage data (sampling frequency is 1 time per minute, including CPU utilization, GPU occupancy rate, storage I / O rate) and task request data (task type, priority, submission time). The training dataset is the resource logs of the past 30 days, the training duration is 12 hours, and the model prediction error (root mean square error) is less than 5%. The resource demand prediction data includes the resource demand for the next 1 hour. For example, it is predicted that 32 CPU cores, 1000GB of GPU video memory, and 2TB of storage space are required during a certain period. As an alternative, the ARIMA model can be used with parameter configuration (p = 2, d = 1, q = 2), which is suitable for scenarios with less resource fluctuations, but the prediction accuracy is slightly reduced to 85%. In actual applications, an AI data center uses the LSTM model to predict peak resource demands with an accuracy of 92%, avoiding resource idleness (the idleness rate drops from 20% to 5%). The prediction data is transmitted to the resource allocation module through the RESTful API, and the response time < 50ms. For different task scenarios, such as AI inference tasks (requiring low latency and high concurrency) and scientific computing tasks (requiring high computing power and long cycles), the model dynamically adjusts the prediction weights (inference task weight 0.6, computing task weight 0.4) to improve the prediction pertinence. In large-scale tests, the prediction data supports the resource allocation of 1000 concurrent tasks, and the task scheduling success rate reaches 98%.

[0028] The reinforcement learning algorithm in step S3 uses the deep deterministic policy gradient algorithm to process the virtual resource pool data and task priority data to generate resource allocation policy data. The prefetch policy processes the high-frequency access data through a distributed cache system to generate optimized data locality data, reducing the cross-node data transmission delay.

[0029] Specifically, the DDPG reinforcement learning algorithm in step S3 is deployed in the PyTorch framework of the server, including an Actor network (for decision-making resource allocation) and a Critic network (for evaluating the allocation effect). The network structure is a 4-layer fully connected neural network (with 256 neurons in each layer). The input data is virtual resource pool data (available CPU cores, GPU video memory, storage capacity) and task priority data (high, medium, and low levels), and the output is resource allocation policy data (such as allocating 4 GPU cores and 16GB of video memory to high-priority tasks). The reward function of the DDPG algorithm is defined as: R = 0.5 * resource utilization rate - 0.3 * energy consumption - 0.2 * latency. After 2000 training iterations, the policy converges, and the resource utilization rate increases by 35%. The prefetch policy caches frequently accessed data (such as AI model weights, training datasets) through a Redis cluster (3 nodes, master-slave replication), with a cache hit rate of 90% and the access latency reduced from 50ms to 10ms. The distributed cache system is configured with 8GB of memory, and the data expiration time is set to 1 hour, supporting high-concurrency access (1000 requests per second). In an actual scenario, an AI training task (Transformer model, 10TB dataset) allocates resources through the DDPG algorithm, and the training time is shortened from 48 hours to 32 hours, with the latency reduced by 40%. As an alternative solution, a genetic algorithm can be used, with a population size of 100 and 50 iterations, and the optimization effect is slightly inferior (the resource utilization rate increases by 30%). In a high-concurrency scenario (1000 inference tasks), the prefetch policy reduces the cross-node data transfer volume by 60% and improves the overall task completion efficiency by 45%.

[0030] Embodiment 2: In the second embodiment of the present invention, the present invention provides a computing power resource scheduling system for an AI multi-service data center. The system includes: A resource management module for obtaining multi-service task request data, processing the task request data based on a time series analysis model, and generating resource demand prediction data; A resource allocation module for converting heterogeneous computing resource data into virtual resource pool data according to the resource demand prediction data, generating priority queue data based on task priority data, and processing the priority queue data through containerization technology for resource isolation and dynamic allocation; A scheduling optimization module for processing virtual resource pool data and task priority data based on a reinforcement learning algorithm to generate resource allocation policy data, and at the same time processing frequently accessed data through a prefetch policy to generate optimized data locality data; A cross-region scheduling module, which is used to generate edge-cloud hierarchical scheduling data according to task latency requirement data and network topology data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined network to generate optimized transmission data.

[0031] Specifically, the resource management module is deployed in the Python environment of the server. It receives task request data (in JSON format, including task ID, type, priority, and resource requirements) based on the Flask framework, processes the data through an LSTM model (3 layers, 128 neurons), and generates resource requirement prediction data (in XML format, including the number of CPU cores, GPU video memory, and storage requirements). The module supports multi-threaded processing, processes 100 requests per second, and the response time is <100ms. The resource allocation module converts physical resources (32-core CPU, 80GB GPU video memory, 10TB storage) into a virtual resource pool through KVM virtualization technology, and generates virtual resource pool data (including resource ID, available quantity, and allocation status). The priority queue is implemented based on Redis, the upper limit of the queue length is 1000, tasks are sorted by priority (high, medium, low), the containerization technology uses Kubernetes, 3 Namespaces are configured to isolate resources for high-priority tasks (such as AI training), and the allocation success rate reaches 95%. The scheduling optimization module runs the DDPG algorithm (implemented by PyTorch, 4-layer neural network), inputs the virtual resource pool data and priority data, and outputs resource allocation strategy data (in JSON format, including task ID and allocated resource quantity). The prefetch strategy caches high-frequency data through Redis, the cache hit rate is 90%, and the latency is reduced by 50%. The cross-region scheduling module manages the InfiniBand network based on the SDN controller (Ryu framework) of the OpenFlow protocol, and generates edge-cloud hierarchical scheduling data (including task ID and allocated node ID). Edge nodes (with a latency of 5ms) process autonomous driving tasks, and the cloud center (with a bandwidth of 100Gbps) processes AI training tasks, and the transmission rate reaches 200Gbps. In a test of a certain data center, the system processes 5000 concurrent tasks, the resource utilization rate reaches 90%, and the task scheduling time is shortened by 35%.

[0032] It also includes: A fault tolerance management module, which is used to process node status data based on the dual-active data center architecture to generate failover data, and combine checkpoint technology to process long-cycle task data to generate task continuity data; A load balancing module, which is used to process task allocation data based on real-time load data to generate load balancing data; The energy efficiency optimization module is used to generate green scheduling data based on task type data and power usage efficiency data, and process processor frequency data and fragmented resource data through dynamic voltage and frequency scaling technology and defragmentation algorithms to generate energy efficiency optimization data.

[0033] Specifically, the fault tolerance management module realizes the coordination of the active-active data center through Zookeeper. The primary data center (in the east, with 1000 servers) and the standby data center (in the west, with 800 servers) synchronize node status data (in JSON format, including CPU load and network status) in real time. Fault detection adopts a heartbeat mechanism (once per second), the fault switching time < 1 second, and the switching success rate is 100%. The checkpoint technology saves the long-term task status through Ceph storage (once every 5 minutes, occupying a bandwidth of 5MB / s), generates task continuity data (including task ID and breakpoint status), and the recovery time < 10 seconds. For example, a gene sequencing task (running for 72 hours) was interrupted due to a node failure, and the module restored the task to the breakpoint, with the lost time < 1 minute. The load balancing module collects real-time load data through Prometheus (sampling frequency of 1 second, including CPU utilization and GPU video memory occupancy), and uses the least connections algorithm to allocate tasks, generating load balancing data (including task ID and target node ID). For example, when node A has a load of 90% and node B has a load of 30%, a new task is allocated to node B, and the overload rate is reduced by 60%. The energy efficiency optimization module generates green scheduling data based on the task type (real-time, non-real-time) and PUE data (PUE of the western node = 1.1). Non-real-time tasks (such as batch rendering) are preferentially allocated to the western node, and the scheduling delay < 100ms. The DVFS technology adjusts the CPU frequency (1.2GHz - 3.8GHz) through the Linux cpufreq tool, and reduces it to 1.5GHz under low load, saving 20% of energy consumption. The defragmentation algorithm scans fragmented resources (1GB video memory, 2 CPU cores), integrates them into allocable units, and reduces the fragmentation rate by 70%. During testing, the module supports 100,000 tasks, the fault recovery rate is 100%, and the energy consumption is reduced by 25%.

[0034] The reinforcement learning algorithm in the scheduling optimization module includes the Deep Deterministic Policy Gradient algorithm, which is used to process virtual resource pool data and task priority data to generate resource allocation policy data. The prefetch policy processes high-frequency access data through a distributed cache system to generate optimized data locality data.

[0035] Specifically, the DDPG algorithm of the scheduling optimization module is deployed in the PyTorch environment. The Actor network (4 layers, 256 neurons) generates resource allocation strategies, and the Critic network evaluates the strategy effects. The inputs are virtual resource pool data (32-core CPU, 80GB GPU video memory) and task priority data (high, medium, low), and the output is resource allocation strategy data (in JSON format, including task ID and allocated resource amount). The reward function is R = 0.5 * resource utilization rate - 0.3 * energy consumption - 0.2 * latency. After 2000 training iterations, the resource utilization rate increases by 35% after convergence. The prefetch policy caches high-frequency data (such as AI model parameters, 10MB per time) through a Redis cluster (3 nodes, 8GB memory), with an expiration time of 1 hour and a hit rate of 90%. The access latency is reduced from 50ms to 10ms. As an alternative, a genetic algorithm (population size 100, 50 iterations) can be used, with a 30% increase in resource utilization rate. In the AI training scenario (Transformer model, 10TB dataset), the module allocates 4 GPU cores, and the training time is shortened from 48 hours to 32 hours, with a 40% reduction in latency. In the high-concurrency scenario (1000 inference tasks), the prefetch policy reduces the cross-node transmission volume by 60%, and the task completion efficiency is increased by 45%. The module supports dynamic adjustment of parameters (such as the weights of the reward function) to adapt to different task requirements (such as latency priority for inference tasks and computing power priority for training tasks), with a 10% improvement in the optimization effect. During testing, the module processes 5000 tasks, with a scheduling success rate of 98% and a 20% reduction in energy consumption.

[0036] It also includes: A monitoring and management module, which is used to obtain resource utilization rate data, store input / output performance data and energy consumption data, process the data based on the data center infrastructure management system, and generate visual resource situation data.

[0037] Specifically, the monitoring and management module collects resource utilization data (CPU utilization 0 - 100%, GPU video memory occupancy 0 - 80GB), storage I / O performance data (throughput 0 - 10GB / s), and energy consumption data (power 0 - 500kW) through Prometheus. The sampling frequency is 1 second, and the data is stored in a time series database (InfluxDB, retained for 30 days). The DCIM system generates visual resource situation data based on Grafana, including dashboards (displaying real-time loads), heat maps (showing node load distributions), and trend charts (predicting resource usage in the next 1 hour). For example, a certain data center monitors 1000 servers. The dashboard shows that the average CPU utilization is 75%, and the heat map identifies high-load nodes (load > 90%). Based on this, the administrator adjusts the task allocation, and the overload rate is reduced by 50%. The module supports integration with the scheduling system through RESTful APIs, with a response time < 50ms and a data processing throughput of 1000 items per second. The alarm mechanism configures thresholds (CPU utilization > 95%, I / O latency > 100ms) and notifies the administrator via email and SMS, with an alarm response time < 5 seconds. In actual applications, the module monitors 100,000 tasks, identifies 100 abnormal nodes, and the fault prediction accuracy rate is 90%. The visual data supports multi-tenant access (based on role permissions). The administrator can view the global situation, and the tenant can view the allocated resources, with an access latency < 200ms. During testing, the module improves the resource management efficiency by 30% and reduces the operation and maintenance costs by 20%.

[0038] Embodiment 3 In the third embodiment of the present invention, based on the same inventive concept, a computer-readable storage medium is proposed. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the computing power resource scheduling method for the AI multi-service data center in the above embodiment.

[0039] Embodiment 4 In the fourth embodiment of the present invention, based on the same inventive concept, a computer device is proposed, including: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the computing power resource scheduling method for the AI multi-service data center in the above embodiment.

[0040] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0041] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A computing power resource scheduling method for an AI multi-service data center, characterized in that It includes the following steps: S1. Obtain multi-service task request data, process the task request data based on a time series analysis model, and generate resource demand prediction data; S2. According to the resource demand prediction data, convert heterogeneous computing resource data into virtual resource pool data, generate priority queue data based on task priority data, and process the priority queue data through containerization technology for resource isolation and dynamic allocation; S3. Process the virtual resource pool data and task priority data based on a reinforcement learning algorithm to generate resource allocation policy data. At the same time, process high-frequency access data through a prefetching policy to generate optimized data locality data to balance resource utilization, energy consumption, and task latency; S4. According to task latency requirement data and network topology data, generate edge-cloud hierarchical scheduling data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined networking to generate optimized transmission data.

2. The computing power resource scheduling method for an AI multi-service data center according to claim 1, wherein It also includes: S5. Process node status data based on a dual-active data center architecture to generate failover data, and process long-cycle task data in combination with checkpoint technology to generate task continuity data; S6. Process task allocation data based on real-time load data to generate load balancing data to avoid local resource overload or idleness; S7. According to task type data and power usage efficiency data, generate green scheduling data, allocate non-real-time task data to low power usage efficiency node data, process processor frequency data through dynamic voltage and frequency scaling technology, and process fragmented resource data in combination with a defragmentation algorithm to generate energy efficiency optimization data.

3. The computing power resource scheduling method for an AI multi-service data center according to claim 1, wherein, The time series analysis model in step S1 includes an autoregressive integrated moving average model or a long short-term memory network model, and the resource demand prediction data includes the predicted usage data of the central processing unit, graphics processing unit, and storage resources within a future period of time.

4. The computing power resource scheduling method for an AI multi-service data center according to claim 1, wherein, The reinforcement learning algorithm in step S3 adopts a deep deterministic policy gradient algorithm to process the virtual resource pool data and task priority data to generate resource allocation policy data. The prefetching policy processes high-frequency access data through a distributed cache system to generate optimized data locality data and reduce cross-node data transmission latency.

5. A computing power resource scheduling system for an AI multi-service data center, characterized in that, For the computing power resource scheduling method for an AI multi-service data center according to any one of claims 1-4, the system includes: A resource management module, configured to obtain multi-service task request data, process the task request data based on a time series analysis model, and generate resource demand prediction data; A resource allocation module, configured to convert heterogeneous computing resource data into virtual resource pool data according to the resource demand prediction data, generate priority queue data based on task priority data, and process the priority queue data through containerization technology for resource isolation and dynamic allocation; A scheduling optimization module, configured to process the virtual resource pool data and task priority data based on a reinforcement learning algorithm to generate resource allocation policy data, and at the same time process high-frequency access data through a prefetching policy to generate optimized data locality data; The cross-region scheduling module is used to generate edge-cloud hierarchical scheduling data according to task latency requirement data and network topology data, allocate low-latency task data to edge node data, allocate large-scale training task data to cloud center data, and process cross-node data transmission through software-defined network to generate optimized transmission data.

6. The computing power resource scheduling system for an AI multi-service data center according to claim 5, wherein It further includes: The fault tolerance management module is used to process node status data based on the active-active data center architecture to generate failover data, and process long-cycle task data in combination with checkpoint technology to generate task continuity data; The load balancing module is used to process task allocation data based on real-time load data to generate load balancing data; The energy efficiency optimization module is used to generate green scheduling data according to task type data and power usage efficiency data, and process processor frequency data and fragmented resource data through dynamic voltage and frequency scaling technology and defragmentation algorithm to generate energy efficiency optimization data.

7. The computing power resource scheduling system for an AI multi-service data center according to claim 5, wherein The reinforcement learning algorithm in the scheduling optimization module includes the Deep Deterministic Policy Gradient algorithm, which is used to process the virtual resource pool data and task priority data to generate resource allocation policy data, and the prefetch policy processes high-frequency access data through a distributed cache system to generate optimized data locality data.

8. The computing power resource scheduling system for an AI multi-service data center according to claim 5, characterized in that, It further includes: The monitoring and management module is used to obtain resource utilization data, storage input / output performance data, and energy consumption data, and process the data based on the data center infrastructure management system to generate visual resource situation data.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the computing power resource scheduling method for the AI multi-service data center according to any one of claims 1 to 4.

10. A readable storage medium, characterized in that, The computer program is stored on the readable storage medium, and when the computer program is executed by the processor, it implements the computing power resource scheduling method for the AI multi-service data center according to any one of claims 1 to 4.

Images