Computing power directional scheduling method and device of intelligent computing center
By dynamically scheduling computing resources in the intelligent computing center, using preset computing resource scheduling algorithms and grayscale release strategies, the problem of fragmentation of computing power resources is solved, resource utilization and operation efficiency are improved, and costs are reduced.
Patent Information
- Application Number
- CN202411295908.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2024-12-06
AI Technical Summary
In the intelligent computing center, the fragmentation of computing power resources leads to low resource utilization, increasing energy consumption and operation and maintenance costs, and traditional scheduling algorithms are difficult to dynamically adjust resource configuration to match task requirements.
By determining the target nodes among all nodes of the intelligent computing center, scheduling tasks to be executed and releasing computing resources on the original node, using the preset computing resource scheduling algorithm to calculate the scheduling weight value to filter the most suitable nodes, and using grayscale Release strategies gradually migrate task traffic to ensure efficient allocation and integration of resources.
It improves the utilization rate of computing power resources, reduces energy consumption and operation and maintenance costs, improves the operation efficiency and response speed of the intelligent computing center, and avoids excessive concentration or idle resources.
Smart Images

Figure CN119088569A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of computing power technology, and in particular, to a method and device for directional scheduling of computing power in an intelligent computing center. Background Art
[0002] With the development of artificial intelligence technology and computing power technology, the concept of intelligent computing center has emerged. "Intelligent computing center" refers to the use of large-scale heterogeneous computing power resources, including general computing power and intelligent computing power, mainly for artificial intelligence applications (such as artificial intelligence deep learning model development, model training and model reasoning and other scenarios) to provide the required computing power, data and algorithms. Intelligent computing center covers facilities, hardware, software, and can provide full-stack capabilities from bottom-level computing power to top-level application enablement.
[0003] Intelligent computing centers usually include a variety of different machine models, each of which has different configurations of CPU (Central Processing Unit), memory, GPU (Graphics Processing Unit) and other hardware computing resources, and the tasks submitted by users require different CPU, memory and GPU resources. The computing resource scheduling of intelligent computing centers mainly relies on some basic scheduling algorithms, such as polling algorithms. As the types and complexity of tasks increase, traditional scheduling algorithms have the following disadvantages:
[0004] In many scenarios, once computing resources are allocated, they are rarely adjusted dynamically based on the task operation status, which causes some computing nodes to be in a high-load state for a long time, while other nodes may be in a low-utilization or idle state. Due to the mismatch between the above-mentioned task requirements and resource supply, resource fragmentation has become a prominent problem. For example, the imbalance in the allocation of CPU resources and memory resources has led to fragmentation of precious GPU resources. If the configuration of computing resources cannot be flexibly adjusted, some computing resources will be idle for a long time.
[0005] In summary, in the existing computing resource scheduling of intelligent computing centers, the fragmentation of computing resources limits the computing power improvement of intelligent computing centers and increases energy consumption and operation and maintenance costs. Summary of the invention
[0006] The embodiments of the present application provide a method and device for directional scheduling of computing power in an intelligent computing center to solve the technical problem in the related art that, in the computing power resource scheduling of the existing intelligent computing center, the fragmentation of computing power resources limits the computing power improvement of the intelligent computing center and increases energy consumption and operation and maintenance costs.
[0007] In order to solve the above technical problems, this application is implemented as follows:
[0008] In a first aspect, an embodiment of the present application provides a method for directional scheduling of computing power in an intelligent computing center, the method comprising:
[0009] Step S1: Determine the tasks to be scheduled;
[0010] Step S2: Determine a target node among all nodes of the intelligent computing center according to a preset computing resource scheduling algorithm;
[0011] Step S3: Schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original nodes.
[0012] Optionally, before step S1, the method further includes:
[0013] Step S01: Obtain the computing resource allocation of historical tasks;
[0014] Step S02: According to the computing power resource allocation situation, the computing power resources on each node of the intelligent computing center are divided according to a preset ratio to obtain different computing power resource scheduling areas; wherein the computing power resources include at least one of the following: central processing unit CPU resources, graphics processing unit GPU resources and memory resources; the computing power resource scheduling areas include: CPU resource scheduling area, GPU resource scheduling area, memory resource scheduling area and operating system reserved area;
[0015] Step S03: Determine the computing power resource requirements of the task to be executed, allocate computing power resources to the task to be executed according to the computing power resource requirements of the task to be executed and the different computing power resource scheduling areas, and execute the task to be executed.
[0016] Optionally, step S1 includes:
[0017] Step S11: every preset period, determine a task to be scheduled from the plurality of tasks to be executed according to the task identifier.
[0018] Optionally, step S2 includes:
[0019] Step S21: Calculate the scheduling weight value of each node of the intelligent computing center according to the computing power resource scheduling algorithm;
[0020] Step S22: Filtering out the minimum scheduling weight value from the scheduling weight values;
[0021] Step S23: Determine the node corresponding to the minimum scheduling weight value as the target node, wherein the minimum scheduling weight value is a positive number.
[0022] Optionally, the step S21 includes:
[0023] Step S211: According to the computing resource scheduling algorithm, the scheduling weight value of each node of the intelligent computing center is calculated by the following formula:
[0024] H = A × (B × F + C × G + D + E / N);
[0025] Wherein, H is a weight value;
[0026] The value of A is a first value or a second value. When the value of A is the first value, it means that the remaining computing power resources of the current node meet the computing power resource requirements of the task to be scheduled; when the value of A is the second value, it means that the remaining computing power resources of the current node do not meet the computing power resource requirements of the task to be scheduled;
[0027] The value of B is a third value or a fourth value. When the value of B is the third value, it represents that there is no task being executed on the current node; when the value of B is the fourth value, it represents that there is a task being executed on the current node;
[0028] F is a first preset threshold, representing the weight of B;
[0029] The C represents the number of the tasks being executed on the current node;
[0030] G is a second preset threshold, representing the weight of C, and G is less than F;
[0031] The D represents the number of remaining CPU resources or GPU resources on the current node;
[0032] The E represents the sequence number of the current node.
[0033] Optionally, step S3 includes:
[0034] Step S31: Locking the target computing power resources on the target node, wherein the target computing power resources are the computing power resources required by the task to be scheduled;
[0035] Step S32: Initializing configuration on the target node based on the target computing resources;
[0036] Step S33: using a grayscale release strategy, gradually migrating the traffic of the task to be scheduled from the original node to the target node;
[0037] Step S34: After the migration is completed, the computing resources occupied by the task to be scheduled on the original node are released.
[0038] In a second aspect, an embodiment of the present application provides a computing power directional scheduling device for an intelligent computing center, the device comprising:
[0039] A determination module is used to determine the tasks to be scheduled;
[0040] The execution module is used to determine the target node among all the nodes of the intelligent computing center according to a preset computing resource scheduling algorithm; schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original node.
[0041] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps of a method for directed scheduling of computing power in an intelligent computing center as described in the first aspect are implemented.
[0042] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for directed scheduling of computing power in an intelligent computing center as described in the first aspect are implemented.
[0043] In a fifth aspect, an embodiment of the present application provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of a method for directed scheduling of computing power in an intelligent computing center as described in the first aspect.
[0044] In the embodiment of the present application, it is first necessary to identify the tasks to be scheduled in order to correctly determine which tasks need to be scheduled or can be scheduled, so as to provide a prerequisite for the subsequent effective resource allocation; then, by using a preset computing power resource scheduling algorithm, the most suitable target node is selected from all nodes of the intelligent computing center, and then the tasks to be scheduled on one or more nodes are scheduled to the target node, and the computing power resources occupied by the tasks to be scheduled on the original node are released, and the reallocation and integration of computing power resources are completed without affecting the execution of the tasks to be scheduled, ensuring that the tasks to be scheduled can be allocated to the most suitable target node, and at the same time, the computing power resources that are no longer needed are released in time, so as to provide space for other tasks. In this way, efficient and balanced computing power resource management can be maintained, excessive concentration or idleness of computing power resources can be avoided, the fragmentation of computing power resources can be effectively reduced, and the computing power resource utilization rate of the intelligent computing center can be significantly improved, thereby improving the operating efficiency and response speed of the entire intelligent computing center, and indirectly reducing energy consumption and operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0046] Figure 1 A flowchart of a method for directional scheduling of computing power in an intelligent computing center provided in an embodiment of the present application;
[0047] Figure 2 A flowchart of a method for directional scheduling of computing power in an intelligent computing center provided in an embodiment of the present application;
[0048] Figure 3 A flowchart of a method for directional scheduling of computing power in an intelligent computing center provided in an embodiment of the present application;
[0049] Figure 4 A schematic diagram of the partitioning of computing resources of an intelligent computing center provided in an embodiment of the present application;
[0050] Figure 5 A structural block diagram of a computing power directional scheduling device for an intelligent computing center provided in an embodiment of the present application;
[0051] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0053] The following is a brief description of the technical terms involved in this application.
[0054] The "computing power" mentioned in this application is the ability of computer equipment or computing / data center to process information. It is the ability of computer hardware and software to work together to perform certain computing needs. It is the computing power to achieve target result output by processing information data. It is a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity. It mainly provides services to society through computing power infrastructure.
[0055] The "computing power (CP)" described in this application is the ability of a data center server to process data and output results. It is a comprehensive indicator to measure the computing power of a data center, including general computing power, supercomputing power and intelligent computing power. The commonly used unit of measurement is the number of floating point operations per second (Floating Point Operations Per Second, FLOPS, where 1EFLOPS=10^18FLOPS). The larger the value, the stronger the comprehensive computing power. According to calculations, 1EFLOPS is approximately the computing power output of 5 Tianhe-2A or 500,000 mainstream server CPUs (Central Processing Unit) or 2 million mainstream notebooks. The calculation formula is: CP=CP 通用 +CP 智能 +CP 超级 In the embodiment of the present application, the computing power adopts the half-precision floating point computing power (FP16) of the graphics card.
[0056] The "carrying capacity" (Network Power, NP) described in this application is a manifestation of the data transmission capability of computing facilities, including comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling. Carrying capacity involves network transmission within and between data centers, and is a comprehensive indicator for measuring network transmission scheduling capabilities. In the embodiment of this application, carrying capacity uses video memory bandwidth. In the embodiment of this application, storage capacity uses video memory bandwidth.
[0057] The "Storage Power" (SP) described in this application refers to the comprehensive capabilities of a data center in terms of data storage capacity, performance, safety and reliability, and green and low-carbon. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and server built-in storage devices. The commonly used unit of measurement for storage capacity is exabyte (EB, 1EB = 2^60bytes), and the commonly used unit of measurement for performance is the number of reads and writes per second per unit capacity (IOPS / TB, Input / Output Operations Per Second / TB). The disaster recovery ratio is an important manifestation of safety and reliability. In the embodiment of this application, storage power uses video memory capacity.
[0058] The "computing power infrastructure" described in this application is a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity. It can realize centralized computing, storage, transmission and application of information, and presents characteristics such as diversity and ubiquity, intelligence and agility, security and reliability, and green and low-carbon.
[0059] The “computing power” mentioned in this application includes general computing power CP 通用、Intelligent computing power CP 智能 And super computing power CP 超级 .
[0060] The "general computing power" mentioned in this application refers to the computing power provided by servers based on CPU chips, which is used to support basic general computing such as cloud computing and edge computing.
[0061] The "intelligent computing power" mentioned in this application refers to the large-scale deployment of intelligent computing centers based on special chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit) for various innovative artificial intelligence applications, such as natural language processing and machine vision.
[0062] The "supercomputing power" mentioned in this application is mainly the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for calculations in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.
[0063] The "intelligent computing center" mentioned in this application refers to a facility that provides the required computing power, data and algorithms for artificial intelligence applications (such as artificial intelligence deep learning model development, model training and model reasoning) by using large-scale heterogeneous computing resources, including general computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.). The intelligent computing center covers facilities, hardware, and software, and can provide full-stack capabilities from bottom-level computing power to top-level application enablement.
[0064] The "computing resources" mentioned in this application refer to technologies and facilities with information computing, transmission, storage and application capabilities required for the development of a digital society, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and supporting and guarantee resources such as wind, fire, water and electricity.
[0065] Please refer to Figure 1 , Figure 1 A method for directional scheduling of computing power in an intelligent computing center provided according to an embodiment of the present application is shown, and the method includes:
[0066] Step S1: Determine the tasks to be scheduled;
[0067] Step S2: Determine the target node among all nodes in the intelligent computing center according to the preset computing resource scheduling algorithm;
[0068] Step S3: Schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original node.
[0069] It should be noted that step S1 includes: step S11: determining a task to be scheduled from a plurality of tasks to be executed according to a task identifier at every preset period.
[0070] In one possible implementation, whether a task in an intelligent computing center can be scheduled (i.e., whether it is a task to be scheduled) can be determined based on a task identifier. The task identifier can be used to indicate whether the task is a stateless task or a stateful task (specifically, when allocating tasks, it can be marked whether the task is a stateless task that occupies resources for a long time and can be migrated, or a stateful task. For stateless tasks, task arrangement can be performed). Stateless tasks do not save task status information, and they can be migrated between different servers or nodes at any time without affecting the operation of the task itself. Common examples include model serving tasks. Stateful tasks are different from stateless tasks. Stateful tasks will maintain and rely on certain task status information during execution, such as data processing, training, and fine-tuning tasks. This status information may include unfinished calculation results, intermediate data, etc. Migrating such tasks requires more resources and time, and may affect the continuity of the tasks. Therefore, the task to be scheduled can be a stateless task.
[0071] When running many tasks on an intelligent computing center, especially on limited hardware resources, the uneven demand for resources (CPU, GPU, etc.) from different tasks will lead to resource fragmentation, especially the fragmentation of precious GPU resources. Resource fragmentation is mainly manifested in some resources being idle or over-concentrated, resulting in low overall resource utilization. By regularly checking (for example, every hour) the use of computing resources, identifying computing resource fragments and sorting them out, resource distribution can be optimized and resource utilization can be improved.
[0072] Specifically, for stateless tasks, since they are easy to migrate, they can be moved to the target node (for example: moving from a machine with high resource usage to a machine with relatively idle resources, or, if the remaining resources of a target node are just enough for the tasks to be scheduled, then in order to ensure the relative concentration and effective utilization of resources, the tasks to be scheduled can be migrated to the target node. The specific method of determining the target node will be described in detail later).
[0073] It should be noted that the fragmentation of GPU resources is mainly caused by the following aspects: Many tasks running on the GPU (such as deep learning model training, graphics rendering, etc.) require not only a large amount of GPU computing power, but also corresponding CPU and memory resources for data preprocessing, data transmission and storage of intermediate results. If the CPU or memory resources are insufficient, even if the GPU is idle, the tasks cannot be started or run inefficiently because they cannot get the necessary support to process or transmit enough data to the GPU; when the computing power resource requirements of different tasks are inconsistent, some computing power resources may be completely exhausted while other computing power resources have a large surplus. For example, a task may consume most of the memory and only use a small amount of GPU computing power resources, which makes other memory-intensive tasks on the same machine unable to start, even though its GPU resources are still a lot of surplus; the efficient computing power of the GPU depends on the ability to quickly and continuously obtain data from the CPU and memory. If the CPU processing speed is insufficient or the memory bandwidth is insufficient, even if the GPU has processing power, it will be idle due to insufficient data supply.
[0074] In a possible implementation, step S2 includes:
[0075] Step S21: Calculate the scheduling weight value of each node of the intelligent computing center according to the computing power resource scheduling algorithm;
[0076] Step S22: Filter out the minimum scheduling weight value from the scheduling weight values;
[0077] Step S23: Determine the node corresponding to the minimum scheduling weight value as the target node, wherein the minimum scheduling weight value is a positive number.
[0078] It should be noted that the scheduling weight value of each node can be calculated based on the preset computing resource scheduling algorithm. These scheduling weight values take into account a variety of factors, such as the current load of the node, historical performance, network bandwidth, storage capacity, etc. (detailed description will be given later). The computing resource scheduling algorithm will dynamically calculate the scheduling weight value of each node based on these parameters to determine which nodes will be identified as target nodes to carry the tasks to be scheduled. After obtaining the scheduling weight values of all nodes, the minimum scheduling weight value will be found out. The minimum weight value is selected to ensure that the tasks to be scheduled can be assigned to the most appropriate nodes, avoid the situation where some nodes are overloaded and other nodes are idle, and realize the reasonable allocation and utilization of computing resources, thereby balancing the load of the entire intelligent computing center and improving operating efficiency.
[0079] Specifically, step S21 includes:
[0080] Step S211: According to the computing power resource scheduling algorithm, the scheduling weight value of each node in the computing intelligent computing center is determined by the following formula:
[0081] H = A × (B × F + C × G + D + E / N);
[0082] Among them, H is the weight value;
[0083] The value of A is the first value or the second value. When the value of A is the first value, it means that the remaining computing resources of the current node meet the computing resource requirements of the task to be scheduled; when the value of A is the second value, it means that the remaining resources of the current node do not meet the computing resource requirements of the task to be scheduled;
[0084] The value of B is the third value or the fourth value. When the value of B is the third value, it means that there is no task being executed on the current node; when the value of B is the fourth value, it means that there is a task being executed on the current node;
[0085] F is the first preset threshold, representing the weight of B;
[0086] C represents the number of tasks being executed on the current node;
[0087] G is the second preset threshold, representing the weight of C, and G is less than F;
[0088] D represents the number of remaining CPU resources or GPU resources on the current node;
[0089] E represents the sequence number of the current node.
[0090] It should be noted that the first value can be 1, and the second value can be 0. When A is 0, it means that the remaining computing power resources of the current node do not meet the computing power resource requirements of the task to be scheduled, and whether the remaining computing power resources of the current node meet the computing power resource requirements of the task to be scheduled is a prerequisite for computing power resource scheduling. If A is 0, H=0 can be obtained according to the formula. According to the limitation in step S23, the minimum weight value should be a positive number, and the node corresponding to the scheduling weight value of H=0 cannot be selected as the target node. When A is 1, it means that the remaining resources of the node at least meet the computing power resource requirements of the task to be scheduled, then it can be temporarily used as an alternative node, and then subsequent priority judgments are performed to screen out the optimal target node. Thus, those nodes that can meet the task requirements can be given priority, thereby improving the execution efficiency and success rate of the task.
[0091] The third value can be 1, and the fourth value can be 0. When B is 1, it means that there is no task being executed on the current node, and when B is 0, it means that there is a task being executed on the current node. In other words, B actually indicates whether the current node is idle. F is its weight, and F can be set to a higher threshold (F can be 100) to emphasize the priority of non-idle nodes, thereby concentrating computing resources and avoiding fragmentation of computing resources.
[0092] C represents the number of tasks being executed, G is its weight, and the value of G is less than F (the value of G can be 10). That is to say, C means that the number of tasks being executed needs to be further considered, but due to the weight setting of G, the influence of the number of tasks being executed is less than the influence of whether the node is idle. Therefore, task allocation can be balanced to prevent node overload.
[0093] D represents the number of remaining CPU resources or GPU resources on the current node. This item is directly added to the weight calculation to reflect the richness of the node's computing resources. The design idea is to try to find nodes with relatively fewer remaining CPU resources or GPU resources while meeting the computing resource requirements of the tasks to be scheduled, so as to concentrate computing resources and avoid fragmentation of computing resources.
[0094] E is the sequence number of the current node, and N is the total number of nodes. Adding E / N to the formula can ensure that nodes with lower sequence numbers are scheduled first.
[0095] Finally, the node with an H value greater than zero and the minimum scheduling weight value is selected as the target node. Or when multiple target nodes need to be selected, the node with an H value greater than zero and the smaller the scheduling weight value, the higher the priority as the target node. Flexible selection can be made according to actual needs.
[0096] In summary, by weighting and reasonably combining various parameters, the efficiency of resource utilization can be maximized by giving priority to nodes that match task requirements; by giving priority to nodes that are executing tasks, computing resources can be concentrated to avoid fragmentation of computing resources; by considering the number of tasks being executed and the remaining resources, overload of certain nodes can be prevented; through reasonable allocation, the intelligent computing center can dynamically allocate computing resources when facing tasks of different types and sizes, improve processing efficiency, reduce waiting time, and thus improve the overall operating performance of the intelligent computing center.
[0097] Please refer to Figure 2 In one possible implementation, Figure 2 As shown, step S3 includes:
[0098] Step S31: Lock the target computing resources on the target node;
[0099] Among them, the target computing resources are the computing resources required for the task to be scheduled;
[0100] Step S32: Initialize configuration on the target node based on the target computing power resources;
[0101] Step S33: Using the grayscale release strategy, gradually migrate the traffic of the task to be scheduled from the original node to the target node;
[0102] Step S34: After the migration is completed, the computing resources occupied by the task to be scheduled on the original node are released.
[0103] It should be noted that the target computing resources on the target node will be identified and locked first. These computing resources may include CPU, GPU, memory, etc. By locking these computing resources, it can be ensured that in the subsequent scheduling process, the target node can meet the performance requirements of the scheduled task, thereby avoiding the scheduling failure or performance degradation of the scheduled task due to insufficient computing resources.
[0104] After locking the target computing power resources, the target node can be initialized and configured, such as setting the necessary software environment, configuring network parameters, installing dependent libraries or services, etc. Through these initialization steps, it can be ensured that the target node can operate normally and provide the expected service when receiving the tasks to be scheduled. After that, the grayscale release strategy can be used to gradually migrate the traffic of the tasks to be scheduled from the original node to the target node. After the migration is completed, the computing power resources occupied by the tasks to be scheduled on the original node can be released. Once all traffic is successfully migrated to the target node and it is confirmed that the target node can stably process these tasks, the intelligent computing center will release the computing resources occupied by the tasks to be scheduled on the original node. This step ensures the effective use of computing power resources, avoids the waste of computing power resources on the original node, and releases computing power for other tasks or services. In this way, the intelligent computing center can dynamically adjust the allocation of computing power resources and optimize overall performance.
[0105] It should be noted that grayscale release is a risk control strategy that gradually transfers traffic from the original node to the target node and monitors the system performance and task operation to ensure that the stability and performance of the entire system will not be affected during the migration process. Specifically, you can set a certain proportion of traffic transfer, observe the response and performance of the target node, and then decide whether to continue to increase the migration traffic.
[0106] Please refer to Figure 3 In one possible implementation, Figure 3 As shown, before step S1, the method further includes:
[0107] Step S01: Obtain the computing resource allocation of historical tasks;
[0108] Step S02: According to the computing power resource allocation situation, the computing power resources on each node of the intelligent computing center are divided according to a preset ratio to obtain different computing power resource scheduling areas;
[0109] The computing power resources include at least one of the following: central processing unit (CPU) resources, graphics processing unit (GPU) resources and memory resources; the computing power resource scheduling area includes: CPU resource scheduling area, GPU resource scheduling area, memory resource scheduling area and operating system reserved area;
[0110] Step S03: Determine the computing resource requirements of the task to be executed, allocate computing resources to the task to be executed according to the computing resource requirements of the task to be executed and different computing resource scheduling areas, and execute the task to be executed.
[0111] The operating system reserved area refers to a portion of computing resources reserved for the operating system in the computing resource allocation. This portion of computing resources is usually not used by other applications or tasks to ensure the stability and performance of the operating system. The existence of the reserved area can prevent the operating system from crashing or performance degradation due to excessive allocation of computing resources.
[0112] In the management and operation of intelligent computing centers, scheduling areas refer to the logical partitioning and division of resources (such as CPU, GPU, memory, etc.) in order to allocate these resources to various computing tasks in a more orderly and efficient manner. Specifically, scheduling areas can be regarded as a virtual division of physical resources, in which each area has a specific resource type and quantity. For example, a scheduling area may be specially configured with a large number of GPU resources and corresponding CPU and memory support to handle graphics processing or deep learning tasks. By dividing resources into different scheduling areas, these resources can be managed and allocated more effectively. The system can dispatch tasks to the most appropriate area according to the specific needs of the task, thereby maximizing resource utilization and task processing efficiency.
[0113] It should be noted that resource partitioning can be performed in advance, that is, all machines in the intelligent computing center are divided into CPU resource scheduling areas and GPU resource scheduling areas according to a fixed ratio. Since the GPU resource scheduling area is very valuable, the GPU resource scheduling area can be given priority (such as Figure 4 shown).
[0114] Resource partitioning refers to dividing the hardware resources (such as CPU, GPU, memory, etc.) of the intelligent computing center into multiple independent areas or groups in a physical or logical manner. Each partition can be managed and optimized independently to meet the resource requirements of specific tasks. This partitioning can be static or dynamic and can be adjusted according to actual load and demand. For example, GPU resources can be allocated to tasks that require a lot of parallel processing power, while CPU and memory resources can be allocated to computing-intensive or data-intensive tasks.
[0115] On the basis of resource partitioning, a resource equal proportion specification strategy can be implemented. Among them, the resource equal proportion specification strategy is: when allocating resources, various resources are configured and allocated according to the preset ratio and specifications. For example, if a task requires 2 CPU cores with 1 GPU, then when configuring the server, resources can be allocated according to this ratio to ensure that each task can get balanced resource support. This strategy helps to reduce the fragmentation problem caused by resource mismatch, so that each task can get all the resources it needs, and improve the overall efficiency of resource use. Alternatively, GPU, CPU and memory can be allocated according to a certain ratio, for example: GPU: CPU: memory = 1:10:20G, that is, for every GPU, 10 CPU cores and 20GB of memory should be equipped at the same time. This ratio can be set based on the typical resource requirements of a specific type of task (such as certain data science or artificial intelligence tasks), aiming to ensure that when the task is executed, the GPU, CPU and memory can achieve optimal work efficiency and coordination (that is, it is equivalent to obtaining the computing power resource allocation of historical tasks, and according to the computing power resource allocation, the computing power resources on each node of the intelligent computing center are divided according to the preset ratio).
[0116] After removing the reserved area, the remaining CPU and memory can be allocated to the CPU task scheduling area, and the CPU and memory are also allocated at a fixed ratio, for example, they can be allocated at a ratio of 1:2G. In other words, after configuring several GPU tasks, the intelligent computing center may find that there are still a considerable amount of CPU and memory resources that are not used. In this case, other CPU-intensive tasks can be scheduled based on the remaining CPU and memory resources according to the ratio of "CPU: memory = 1:2G" to effectively configure and use these remaining resources.
[0117] It should be noted that in a possible implementation, after executing the resource partitioning and resource equal proportion specification strategy, a computing power resource regular arrangement plan can also be executed, that is, in a specific application scenario, Figure 3 The solution shown in the figure is then executed Figure 1The solution for regular arrangement of computing resources is shown in the figure. Specifically, the use of resource partitioning and resource proportional specification strategies can avoid the generation of resource fragmentation due to unreasonable resource settings during computing resource allocation to the greatest extent, which in turn causes other nodes to be unable to be scheduled. However, after a period of operation, resource fragmentation will be generated to varying degrees on multiple machines. At this time, it is necessary to adopt a regular arrangement strategy to arrange computing resources to avoid the fragmentation of computing resources, especially the precious GPU computing resources, and improve the utilization rate of computing resources.
[0118] In summary, in the embodiments of the present application, it is first necessary to identify the tasks to be scheduled in order to correctly determine which tasks need to be scheduled or can be scheduled, so as to provide a prerequisite for the subsequent effective resource allocation; then, by using a preset computing power resource scheduling algorithm, the most suitable target node is selected from all the nodes of the intelligent computing center, and then the tasks to be scheduled on one or more nodes are scheduled to the target node, and the computing power resources occupied by the tasks to be scheduled on the original node are released. Without affecting the execution of the tasks to be scheduled, the computing power resources are redistributed and integrated to ensure that the tasks to be scheduled can be allocated to the most suitable target node, and at the same time, computing power resources that are no longer needed are released in time to provide space for other tasks, so as to maintain efficient and balanced computing power resource management, avoid excessive concentration or idleness of computing power resources, effectively reduce the fragmentation of computing power resources, and significantly improve the computing power resource utilization rate of the intelligent computing center, thereby improving the operating efficiency and response speed of the entire intelligent computing center, and indirectly reducing energy consumption and operation and maintenance costs.
[0119] Please refer to Figure 5 The embodiment of the present application further provides a computing power directional scheduling device 50 of an intelligent computing center, and the device 50 includes:
[0120] A determination module 51 is used to determine the tasks to be scheduled;
[0121] The execution module 52 is used to determine the target node among all the nodes of the intelligent computing center according to the preset computing resource scheduling algorithm; schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original node.
[0122] In one possible implementation, the execution module 52 is also used to obtain the computing power resource allocation of historical tasks; according to the computing power resource allocation, the computing power resources on each node of the intelligent computing center are divided according to a preset ratio to obtain different computing power resource scheduling areas; wherein the computing power resources include at least one of the following: central processing unit CPU resources, graphics processing unit GPU resources and memory resources; the computing power resource scheduling area includes: CPU resource scheduling area, GPU resource scheduling area, memory resource scheduling area and operating system reserved area; determine the computing power resource requirements of the task to be executed, allocate computing power resources to the task to be executed according to the computing power resource requirements of the task to be executed and the different computing power resource scheduling areas, and execute the task to be executed.
[0123] In a possible implementation, the determination module 51 is further configured to determine the to-be-scheduled task from a plurality of to-be-executed tasks according to the task identifier at every preset period.
[0124] In one possible implementation, the execution module 52 is also used to calculate the scheduling weight value of each node of the intelligent computing center according to the computing power resource scheduling algorithm; among the scheduling weight values, the minimum scheduling weight value is screened out; and the node corresponding to the minimum scheduling weight value is determined as the target node, wherein the minimum scheduling weight value is a positive number.
[0125] In a possible implementation, the execution module 52 is further configured to calculate the scheduling weight value of each node of the intelligent computing center according to the computing power resource scheduling algorithm by using the following formula:
[0126] H = A × (B × F + C × G + D + E / N);
[0127] Among them, H is the weight value;
[0128] The value of A is the first value or the second value. When the value of A is the first value, it means that the remaining computing resources of the current node meet the computing resource requirements of the task to be scheduled; when the value of A is the second value, it means that the remaining resources of the current node do not meet the computing resource requirements of the task to be scheduled;
[0129] The value of B is the third value or the fourth value. When the value of B is the third value, it means that there is no task being executed on the current node; when the value of B is the fourth value, it means that there is a task being executed on the current node;
[0130] F is the first preset threshold, representing the weight of B;
[0131] C represents the number of tasks being executed on the current node;
[0132] G is the second preset threshold, representing the weight of C, and G is less than F;
[0133] D represents the number of remaining CPU resources or GPU resources on the current node;
[0134] E represents the sequence number of the current node.
[0135] In one possible implementation, the execution module 52 is also used to lock the target computing power resources on the target node, wherein the target computing power resources are the computing power resources required for the task to be scheduled; perform initialization configuration on the target node based on the target computing power resources; utilize the grayscale release strategy to gradually migrate the traffic of the task to be scheduled from the original node to the target node; after the migration is completed, release the computing power resources occupied by the task to be scheduled on the original node.
[0136] In summary, in the embodiments of the present application, it is first necessary to identify the tasks to be scheduled in order to correctly determine which tasks need to be scheduled or can be scheduled, so as to provide a prerequisite for the subsequent effective resource allocation; then, by using a preset computing power resource scheduling algorithm, the most suitable target node is selected from all the nodes of the intelligent computing center, and then the tasks to be scheduled on one or more nodes are scheduled to the target node, and the computing power resources occupied by the tasks to be scheduled on the original node are released. Without affecting the execution of the tasks to be scheduled, the computing power resources are redistributed and integrated to ensure that the tasks to be scheduled can be allocated to the most suitable target node, and at the same time, computing power resources that are no longer needed are released in time to provide space for other tasks, so as to maintain efficient and balanced computing power resource management, avoid excessive concentration or idleness of computing power resources, effectively reduce the fragmentation of computing power resources, and significantly improve the computing power resource utilization rate of the intelligent computing center, thereby improving the operating efficiency and response speed of the entire intelligent computing center, and indirectly reducing energy consumption and operation and maintenance costs.
[0137] Please refer to Figure 6 The embodiment of the present application also provides an electronic device 60, including a processor 61, a memory 62, and a computer program stored in the memory 62 and executable on the processor 61. When the computer program is executed by the processor 61, the various processes of the embodiment of the computing power directional scheduling method of the above-mentioned intelligent computing center are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.
[0138] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, each process of the above-mentioned intelligent computing center computing power directional scheduling method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. Among them, the computer-readable storage medium is such as a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.
[0139] The embodiment of the present application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes of the embodiment of the computing power directional scheduling method of the above-mentioned intelligent computing center are implemented, and the same technical effect can be achieved. To avoid repetition, they will not be repeated here.
[0140] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.
[0141] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.
[0142] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.
Claims
1. A method for directional scheduling of computing power in an intelligent computing center, characterized in that: The method comprises: Step S1: Determine the tasks to be scheduled; Step S2: Determine a target node among all nodes of the intelligent computing center according to a preset computing resource scheduling algorithm; Step S3: Schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original nodes.
2. The method according to claim 1, characterized in that Before step S1, the method further includes: Step S01: Obtain the computing resource allocation of historical tasks; Step S02: According to the computing power resource allocation situation, the computing power resources on each node of the intelligent computing center are divided according to a preset ratio to obtain different computing power resource scheduling areas; wherein the computing power resources include at least one of the following: central processing unit CPU resources, graphics processing unit GPU resources and memory resources; the computing power resource scheduling areas include: CPU resource scheduling area, GPU resource scheduling area, memory resource scheduling area and operating system reserved area; Step S03: Determine the computing power resource requirements of the task to be executed, allocate computing power resources to the task to be executed according to the computing power resource requirements of the task to be executed and the different computing power resource scheduling areas, and execute the task to be executed.
3. The method according to claim 2, characterized in that Step S1 includes: Step S11: every preset period, determine a task to be scheduled from the plurality of tasks to be executed according to the task identifier.
4. The method according to claim 1, characterized in that: Step S2 includes: Step S21: Calculate the scheduling weight value of each node of the intelligent computing center according to the computing power resource scheduling algorithm; Step S22: Filtering out the minimum scheduling weight value from the scheduling weight values; Step S23: Determine the node corresponding to the minimum scheduling weight value as the target node, wherein the minimum scheduling weight value is a positive number.
5. The method according to claim 4, characterized in that The step S21 comprises: Step S211: According to the computing resource scheduling algorithm, the scheduling weight value of each node of the intelligent computing center is calculated by the following formula: H = A × (B × F + C × G + D + E / N); Wherein, H is a weight value; The value of A is a first value or a second value. When the value of A is the first value, it means that the remaining computing power resources of the current node meet the computing power resource requirements of the task to be scheduled; when the value of A is the second value, it means that the remaining computing power resources of the current node do not meet the computing power resource requirements of the task to be scheduled; The value of B is a third value or a fourth value. When the value of B is the third value, it represents that there is no task being executed on the current node; when the value of B is the fourth value, it represents that there is a task being executed on the current node; F is a first preset threshold, representing the weight of B; The C represents the number of the tasks being executed on the current node; G is a second preset threshold, representing the weight of C, and G is less than F; The D represents the number of remaining CPU resources or GPU resources on the current node; The E represents the sequence number of the current node.
6. The method according to any one of claims 1 to 5, characterized in that The step S3 comprises: Step S31: Locking the target computing power resources on the target node, wherein the target computing power resources are the computing power resources required by the task to be scheduled; Step S32: Initializing configuration on the target node based on the target computing resources; Step S33: using a grayscale release strategy, gradually migrating the traffic of the task to be scheduled from the original node to the target node; Step S34: After the migration is completed, the computing resources occupied by the task to be scheduled on the original node are released.
7. A computing power directional scheduling device for an intelligent computing center, characterized in that: The device comprises: A determination module is used to determine the tasks to be scheduled; The execution module is used to determine the target node among all the nodes of the intelligent computing center according to a preset computing resource scheduling algorithm; schedule the tasks to be scheduled on one or more nodes to the target node, and release the computing resources occupied by the tasks to be scheduled on the original node.
8. An electronic device, characterized in that: include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps of a method for directional scheduling of computing power in an intelligent computing center as described in any one of claims 1 to 6 are implemented.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for directional scheduling of computing power in an intelligent computing center as described in any one of claims 1 to 6.
10. A computer program product, characterized in that It includes computer instructions, which, when executed by a processor, implement the steps of a computing power directional scheduling method for an intelligent computing center as described in any one of claims 1 to 6.
Citation Information
Cited By
Intelligent computing power scheduling method and device for intelligent computing center of common computing power
CN120029739A
Resource allocation method and apparatus, and electronic device
CN120295796A
Resource allocation method, device, and electronic equipment
CN120295796B
