Computing power resource management method of intelligent computing center and related device

Abstract

The invention provides a computing power resource management method of an intelligent computing center and a related device. The method comprises the following steps: S1, predicting a computing power demand of the intelligent computing center at a future moment by adopting a computing power demand prediction large language model deployed in the intelligent computing center; s2, determining a computing power resource scheduling scheme from the current moment to the future moment according to the computing power demand; and S3, executing the computing power resource scheduling scheme until the future moment. According to the method, the computing power resources can be managed in time, and the management efficiency is improved.

Description

Technical Field

[0001] Embodiments of the present invention relate to the technical field of computing power infrastructure, and in particular, to a computing power resource management method and related devices for 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] Computing resources refer to technologies and facilities with the ability to calculate, transmit, store and apply information required for the development of a digital society, including but not limited to computing resources such as CPU (Central Processing Unit) and GPU (Graphics Processing Unit), 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.

[0004] The existing methods for managing computing resources in intelligent computing centers are mostly reactive management performed after actual resource demand arises. This leads to high management lag and low management efficiency. Summary of the invention

[0005] The embodiments of the present invention provide a computing resource management method and related devices for an intelligent computing center to solve the problem that the existing computing resource management methods for intelligent computing centers are mostly reactive management performed after actual resource demand arises, with high management lag and low management efficiency.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0007] In a first aspect, an embodiment of the present invention provides a computing resource management method for an intelligent computing center, including:

[0008] Step S1: using a large language model for predicting computing power requirements deployed in an intelligent computing center to predict the computing power requirements of the intelligent computing center at a future time;

[0009] Step S2: Determine a computing resource scheduling plan from the current moment to the future moment according to the computing power demand;

[0010] Step S3: Execute the computing resource scheduling plan until the future moment.

[0011] Optionally, step S2 includes:

[0012] Step S21: inputting the computing power demand into a resource scheduling large language model deployed in an intelligent computing center to determine the computing power resource scheduling solution;

[0013] The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

[0014] Optionally, the resource scheduling large language model is trained according to at least one of the following constraints: resource capacity constraint, task requirement constraint, quota constraint, and logical isolation constraint.

[0015] Optionally, before step S1, the following steps are included:

[0016] Step S01: receiving a scheduling request sent by a user through an interactive terminal, wherein the scheduling request includes user identity information; based on the user identity information, verifying whether the user has access rights to access computing resources of an intelligent computing center, and obtaining a verification result;

[0017] Step S02: if the verification result indicates that the user has the access right, determine the computing resource quota of the user in the intelligent computing center, and execute the steps S1, S2 and S3 based on the computing resources of the computing resource quota;

[0018] Step S03: If the verification result indicates that the user does not have the access authority, step S1 is not executed.

[0019] Optionally, the computing power resource quota is determined according to a quota constraint condition, and the expression of the quota constraint condition is as follows:

[0020]

[0021] Among them, x u,p,r The allocation of resource type r for user u on project p; The resource quota limit for user u in project p; For all users u.

[0022] In a second aspect, an embodiment of the present invention provides a computing resource management device for an intelligent computing center, including:

[0023] A prediction module, used to use a large language model for predicting computing power requirements deployed in an intelligent computing center to predict the computing power requirements of the intelligent computing center at future moments;

[0024] A determination module, used to determine a computing power resource scheduling plan from the current moment to the future moment according to the computing power demand;

[0025] An execution module is used to execute the computing resource scheduling plan until the future moment.

[0026] Optionally, the determination module is further used to input the computing power demand into a resource scheduling large language model deployed in the intelligent computing center to determine the computing power resource scheduling solution;

[0027] The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

[0028] In a third aspect, an embodiment of the present invention provides an electronic device comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps in the computing power resource management method of an intelligent computing center as described in any one of the first aspects.

[0029] In a fourth aspect, an embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps in the computing power resource management method of an intelligent computing center as described in any one of the first aspects are implemented.

[0030] In a fifth aspect, an embodiment of the present invention provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of the computing power resource management method of an intelligent computing center as described in any one of the first aspects.

[0031] In an embodiment of the present invention, step S1 is performed: a large language model for predicting computing power demand deployed in the intelligent computing center is used to predict the computing power demand of the intelligent computing center at future moments; step S2 is performed: a computing power resource scheduling plan from the current moment to the future moment is determined based on the computing power demand; step S3 is performed: the computing power resource scheduling plan is executed until the future moment. The embodiment of the present invention performs advanced computing power resource scheduling plan planning based on predicted demand, which solves the problem of high management lag of existing management methods, can realize timely management of computing power resources, and improves management efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] 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 invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0033] Figure 1 A flowchart of a computing resource management method for an intelligent computing center according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram comparing the actual computing power demand with the predicted computing power demand;

[0035] Figure 3 This is a principle block diagram of a computing resource management device for an intelligent computing center according to an embodiment of the present invention;

[0036] Figure 4 The figure is a functional block diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The terms "first", "second", etc. of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of one type, and the number of objects is not limited, for example, the first object can be one or more. In addition, "or" in the present application represents at least one of the connected objects. For example, "A or B" covers three schemes, namely, Scheme 1: including A but not including B; Scheme 2: including B but not including A; Scheme 3: including both A and B. The character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0039] In the technical solution of the present disclosure, words such as “connect”, “couple” or “connected” are not limited to physical or mechanical connections, but may include electrical connections.

[0040] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0041] First, the technical terms involved in the present invention are briefly explained below.

[0042] The "computing power" mentioned in the present invention 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 execute certain computing requirements. 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 the society through computing power infrastructure.

[0043] The "computing power" (CP) described in the present invention 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, super computing power and intelligent computing power. The commonly used unit of measurement is the number of floating point operations performed per second (FLOPS: Floating Point Operations Per Second, 1EFLOPS = 10^18FLOPS). The larger the value, the stronger the comprehensive computing power. According to calculations, 1EFLOPS is about the computing power output of 5 Tianhe-2A or 500,000 mainstream server CPUs or 2 million mainstream notebooks. The calculation formula is: CP = CP general + CP intelligent + CP super.

[0044] The "Network Power" (NP) described in the present invention 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, etc. The 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 the present invention, the carrying capacity uses the memory bandwidth.

[0045] The "Storage Power" (SP) described in the present invention is the comprehensive ability 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.

[0046] The "computing power infrastructure" described in the present invention 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.

[0047] The “computing power” mentioned in the present invention includes general computing power, intelligent computing power and super computing power.

[0048] The “general computing power” mentioned in the present invention refers to the computing power provided by servers based on CPU (Central Processing Unit) chips, which is used to support basic general computing such as cloud computing and edge computing.

[0049] The "intelligent computing power" described in the present invention is a computing platform for large-scale deployment of special chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit) for various innovative applications of artificial intelligence, such as natural language processing and machine vision.

[0050] The "super computing power" mentioned in the present invention 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.

[0051] The "intelligent computing center" described in the present invention 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 scenarios) by using large-scale heterogeneous computing power resources, including general computing power (CPU: Central Processing Unit) and intelligent computing power (GPU: Graphics Processing Unit, FPGA: Field Programmable Gate Array, ASIC: Application Specific Integrated Circuit, 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.

[0052] The "computing resources" mentioned in the present invention refer to the technologies and facilities with information calculation, transmission, storage and application capabilities required for the development of the digital society, including but not limited to computing resources such as CPU (Central Processing Unit), GPU (Graphics Processing Unit), 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, as well as supporting and guarantee resources such as wind, fire, water and electricity.

[0053] The "Multimodal Large Models" mentioned in the present invention refer to models that combine multimodal information such as text, images, videos, audio, etc. for training, including but not limited to multimodal large language models.

[0054] The "large language model" described in the present invention refers to a large language model (LLM), which is a language model with a large parameter scale, designed to understand and generate human language, trained with a large amount of text data, and can perform a wide range of tasks including text summarization, translation, sentiment analysis, etc. Among them, the large language model can be a large language model such as GPT4, GLM-4 or Llama-3.

[0055] The embodiment of the present invention provides a computing resource management method for an intelligent computing center, see Figure 1 As shown, Figure 1 The flowchart of the computing resource management method of the intelligent computing center according to the embodiment of the present invention includes:

[0056] Step S1: using the computing power demand prediction large language model deployed in the intelligent computing center to predict the computing power demand of the intelligent computing center at a future time;

[0057] Step S2: Determine the computing resource scheduling plan from the current moment to the future moment according to the computing power demand;

[0058] Step S3: Execute the computing resource scheduling plan until a future time.

[0059] It should be noted that the large language model for computing power demand prediction is deployed in the intelligent computing center, which means that the large language model for computing power demand prediction is migrated to the intelligent computing center (that is, the production environment) so that the large language model for computing power demand can perform actual reasoning or prediction.

[0060] In the embodiment of the present invention, the future time can be specifically set by the user according to the actual computing resource management needs.

[0061] It can be understood that the embodiment of the present invention can be set to be executed once every preset time period, and the future moment of each execution is variable. For example, the preset time period is 1 hour, that is, the embodiment of the present invention is executed once every 1 hour, and the time difference between the future moments of two adjacent executions of the embodiment of the present invention is also 1 hour. Specifically, the embodiment of the present invention is executed once every 1 hour. If the future moment of the previous execution can be 8 am, the future moment of this execution is 9 am on the same day, and so on; by matching the execution interval (that is, the preset time period) of the embodiment of the present invention with the future moment in the execution process, hourly computing resource management can be achieved.

[0062] For another example, in order to improve the management accuracy of computing power resources, the preset time period is 0.5 hours, that is, the embodiment of the present invention is executed once every 0.5 hours, and the time difference between the future moments when the embodiment of the present invention is executed twice is also 0.5 hours. Thus, the embodiment of the present invention is executed once every 0.5 hours. If the future moment of the previous execution can be 8:00 a.m., the future moment of this execution is 8:30 a.m. of the same day, and so on. By matching the execution interval (that is, the preset time period) of the embodiment of the present invention with the future moment in the execution process, half-hour computing power resource management can be achieved, thereby improving the management accuracy of computing power resources.

[0063] In practical applications, a time margin can also be set to ensure efficient use of computing resources. For example, according to the historical execution of the embodiment of the present invention, the time margin is determined to be 20 minutes. Accordingly, the embodiment of the present invention is executed every 40 minutes. If the future time can be 8 am during the previous execution, the future time can be 9 am on the same day during this execution, and a 20-minute time margin is reserved for the execution of the embodiment of the present invention. By analogy, there is no window period between the execution of the computing resource scheduling scheme in the previous step S3 and the execution of the computing resource scheduling scheme in the next step S3, thereby ensuring efficient use of the computing resources of the intelligent computing center.

[0064] It should be noted that in the embodiment of the present invention, the computing power resource scheduling scheme includes scheduling of at least one of the following computing power resources: computing resources such as CPU (Central Processing Unit) and GPU (Graphics Processing Unit), 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] In an embodiment of the present invention, step S1 is performed: a large language model for predicting computing power demand deployed in the intelligent computing center is used to predict the computing power demand of the intelligent computing center at future moments; step S2 is performed: a computing power resource scheduling plan from the current moment to the future moment is determined based on the computing power demand; step S3 is performed: the computing power resource scheduling plan is executed until the future moment. The embodiment of the present invention performs advanced computing power resource scheduling plan planning based on predicted demand, which solves the problem of high management lag of existing management methods, can realize timely management of computing power resources, and improves management efficiency.

[0066] In some embodiments of the present invention, optionally, step S2 includes:

[0067] Step S21: inputting the computing power demand into a resource scheduling large language model deployed in an intelligent computing center to determine the computing power resource scheduling solution;

[0068] The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

[0069] It should be noted that the resource scheduling large language model is deployed in the intelligent computing center, which means that the resource scheduling large language model is migrated to the intelligent computing center (that is, the production environment) so that the resource scheduling large language model can perform actual reasoning or prediction.

[0070] In the embodiment of the present invention, the first objective function is used to train a resource scheduling language model that minimizes resource usage cost and energy consumption. Specifically, the expression of the first objective function is as follows:

[0071]

[0072] Among them, x i,j is the computational amount allocated to task j by resource i; C i,j is the cost of executing task j on resource i; E i (x i ) is the resource i with an allocation of x i energy consumption function when the system is running; α is the energy consumption weight coefficient, which is used to balance the cost and energy consumption; M is the type or quantity of resources, which refers to the number of different types of resources or resource pools, such as different servers, storage devices or network devices; N is the number of tasks or jobs, which refers to the various computing tasks or jobs that need to be scheduled.

[0073] In an embodiment of the present invention, through step S21: the computing power demand is input into a resource scheduling large language model deployed in an intelligent computing center to determine the computing power resource scheduling plan; wherein the resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption, and can determine a computing power resource scheduling plan that minimizes resource usage cost and minimizes energy consumption according to the predicted computing power demand, thereby improving the utilization rate of computing power resources.

[0074] In some embodiments of the present invention, optionally, the resource scheduling large language model is trained according to at least one of the following constraints: resource capacity constraint, task requirement constraint, quota constraint, and logical isolation constraint.

[0075] The following is an explanation with reference to specific examples.

[0076] 1. The expression of resource capacity constraint is as follows:

[0077]

[0078] Among them, x i,jThe amount of computation allocated to task j by resource i; is the maximum capacity of resource i.

[0079] 2. The expression of task requirement constraint is as follows:

[0080]

[0081] Among them, x i,j The amount of computation allocated to task j by resource i; is the minimum requirement for task j.

[0082] 3. The quota constraint is expressed as follows:

[0083]

[0084] Among them, x i,j,u is the computational amount that user u allocates to task j for resource i; The resource quota upper limit of user u;

[0085] 4. The expression of the logical isolation constraint is as follows:

[0086]

[0087] Among them, x i,j,u R is the amount of computation that user u allocates to resource i for task j; u A collection of resources accessible to user u.

[0088] In the expressions of the above four constraints, M is the type or quantity of resources, which refers to the number of different types of resources or resource pools, such as different servers, storage devices or network devices; N is the number of tasks or jobs, which refers to the various computing tasks or jobs that need to be scheduled.

[0089] It should be noted that in some optional embodiments of the present invention, when the resource scheduling large language model is trained according to resource capacity constraints, task requirement constraints, quota constraints, and logical isolation constraints, the resource scheduling large language model can output a high-accuracy computing resource scheduling plan.

[0090] In some embodiments of the present invention, optionally, before step S1, the steps include:

[0091] Step S01: receiving a scheduling request sent by a user through an interactive terminal, wherein the scheduling request includes user identity information; based on the user identity information, verifying whether the user has access rights to access computing resources of an intelligent computing center, and obtaining a verification result;

[0092] Step S02: if the verification result indicates that the user has the access right, determine the computing resource quota of the user in the intelligent computing center, and execute the steps S1, S2 and S3 based on the computing resources of the computing resource quota;

[0093] Step S03: If the verification result indicates that the user does not have the access authority, step S1 is not executed.

[0094] It should be noted that, through step S01, the user who initiates the computing resource scheduling request can be authenticated to ensure the security of the intelligent computing center.

[0095] Furthermore, for users with access rights, step S02 is executed: if the verification result indicates that the user has the access rights, the computing power resource quota of the user in the intelligent computing center is determined, and the computing power resources based on the computing power resource quota are executed, thereby achieving the isolation of computing power resources of different users in the intelligent computing center and ensuring the security of the intelligent computing center.

[0096] The following is an explanation with reference to specific examples.

[0097] Logical isolation: Build access control models based on user, project and other dimensions to ensure logical isolation of resources. Use permission matrix or access control list (ACL) to define user access rights to resources.

[0098] The expression of the authority matrix A(u,p,r) is as follows:

[0099]

[0100] In some embodiments of the present invention, optionally, the computing power resource quota is determined according to a quota constraint condition, and the expression of the quota constraint condition is as follows:

[0101]

[0102] Among them, x u,p,r is the allocation of resource type r used by user u on project p. Specifically, x u,p,r It represents the actual allocation value of a certain type of resource r (such as CPU, GPU, storage) by user u in a specific task or project p. The unit is computing time, storage capacity or other measurement units depending on the resource type.

[0103] is the resource quota upper limit of user u in project p. Specifically, The maximum amount of available resources that the system allocates to user u on project p. It is usually pre-set to prevent resource abuse or exceeding the allowed range. Unit: same as x u,p,r Consistency;

[0104] Specifically, the quota constraints apply to all users in the system, ensuring that the total resource allocation of each user does not exceed the maximum quota on its project.

[0105] The following is a description with reference to specific embodiments.

[0106] The training of the large language model for computing power demand prediction includes:

[0107] A. Data preparation:

[0108] 200 days of simulated resource demand data were generated and normalized.

[0109] Use the look_back parameter to divide the time series into input and output data for training. look_back=5 means using the data of the first 5 time steps to predict the demand for the next time step.

[0110] B. LSTM model:

[0111] A two-layer LSTM structure is used, with return_sequences set to True and False respectively. This is a common deep learning time series modeling method.

[0112] The model uses the Adam optimizer and is trained by minimizing the mean squared error (MSE).

[0113] C. Training and testing:

[0114] The dataset is split into 80% training set and 20% testing set.

[0115] After training the model, predictions are made on the test set and the results are denormalized to the original data scale.

[0116] D. Chart display:

[0117] Use matplotlib to draw a comparison chart of actual values ​​and predicted values, see Figure 2 As shown, specifically, the X-axis (Date): represents time, in dates. The figure shows the date series from June 15, 2023 to July 15, 2023. It represents the time series forecast of resource demand. The model predicts future resource demand based on data from the past period of time.

[0118] Y-axis (Resource Demand): represents the resource demand in units of a specific demand metric (such as CPU, memory, or other computing resources). The Y-axis ranges from approximately 35 to 60, representing daily resource demand fluctuations.

[0119] Blue line (Actual Resource Demand):

[0120] The blue line represents the actual resource demand, and the data is generated by simulating historical data. Each point corresponds to the actual resource demand for one day, and the blue line and dots represent the fluctuation of historical data.

[0121] from Figure 2 It can be seen that the actual resource demand is highly volatile, especially in certain periods (such as mid-June and early July), when resource demand drops or rises sharply.

[0122] Orange line (Predicted Resource Demand):

[0123] The orange line indicates the resource demand predicted by the model. The LSTM model makes predictions based on data from the past period of time, and these predictions are represented by orange lines and "X" marks.

[0124] The forecast results show a certain smooth trend. Although it can generally follow the changes in actual resource demand, it shows certain lags and errors during periods of sharp fluctuations in resource demand (such as the decline in mid-June or the fluctuations in early July).

[0125] Performance of model predictions:

[0126] Trend Capture: The LSTM model is able to capture the overall changing trend of resource requirements. Figure 2 In the figure, the orange line changes in the same direction as the blue line, indicating that the model is able to identify the overall upward and downward trends in demand and can achieve accurate forecasts of demand.

[0127] Forecast Error: However, the forecast value cannot perfectly follow the actual demand, especially during periods of rapid demand changes. For example, during certain periods (such as around June 15), the sharp fluctuations in actual demand were not accurately predicted by the model, and the forecast curve was much smoother than the actual data.

[0128] The resource scheduling large language model implements scheduling based on performance normalization and collaborative efficiency.

[0129] Performance normalization: Establish performance normalization models for different types of computing resources to facilitate comparison and scheduling.

[0130] Performance normalization coefficient γi:

[0131]

[0132] Among them, P i is the performance indicator of resource i, such as computing power, memory bandwidth, etc.

[0133] Resource collections can be used to train models, especially for large language models (such as LLaMA), which require the coordination of a large amount of computing resources for training. During the training phase, resources in the resource collection are dynamically allocated according to demand to meet the requirements of computing, storage, and data processing.

[0134] Resource Scheduling After the large language model training is completed, the resource collection can still be called for model deployment and operation. In the deployment phase, the role of the resource collection is to support model reasoning (such as real-time response to user requests) or perform continuous model optimization tasks. The system uses an intelligent scheduling mechanism to dynamically allocate appropriate computing resources from the resource collection to support the efficient operation of the model.

[0135] Collaborative efficiency: The execution efficiency of computing tasks on different resources may be different. Define the collaborative efficiency coefficient φ i,j :

[0136]

[0137] Collaborative optimization: In the resource allocation optimization model, collaborative efficiency is introduced to optimize the collaborative work of heterogeneous resources.

[0138] In the present invention, in the deployment and application of the large language model for computing power demand prediction and the large language model for resource scheduling, especially in the execution process of step S1 and step S21 of the embodiment of the present invention, the execution subject of the embodiment of the present invention (hereinafter referred to as the management system, or system) also performs actual reasoning or prediction of the above-mentioned large language model in the intelligent computing center to specify and optimize the physical service cluster, thereby improving resource utilization efficiency.

[0139] Physical service cluster assignment and optimization, that is, by fully considering the physical location of computing resources and network topology, improve resource scheduling efficiency and reduce communication costs. This optimization strategy is not only applicable to resource scheduling within the same data center, but can also be extended to resource management in multiple data centers or multi-cloud environments.

[0140] Physical service cluster designation and optimization:

[0141] a) Definition and allocation of physical service clusters

[0142] In the management of heterogeneous computing resources, a physical service cluster refers to a group of computing resource units with similar physical locations or topological structures, such as physical servers, GPU servers, or computing nodes. When creating a computing resource pool, the system supports administrators or intelligent scheduling systems to specify specific physical service clusters based on actual needs. By reasonably dividing and specifying physical clusters, the system can effectively reduce the communication overhead and network latency between different physical nodes.

[0143] For example, when resources need to be allocated for a computing task, the system can selectively schedule the task to the nearest physical cluster, thereby avoiding the additional delay caused by cross-cluster and cross-network data transmission. This resource pool management mode based on physical service clusters enables the system to have more efficient resource utilization and scheduling capabilities when processing tasks with high computing volume and low latency requirements.

[0144] b) Consider physical topology and network latency

[0145] Physical topology refers to the layout of physical connections between computing resources within or across data centers. For example, some computing nodes may be in the same rack or computer room, connected by high-speed switches, with extremely low communication latency between them; while other nodes may be located in different computer rooms or even different geographical regions, and the network latency between them may be significantly increased.

[0146] In this context, the system needs to fully consider the physical location and network topology of computing resources when allocating resources. Traditional resource scheduling algorithms usually only focus on the availability of computing resources, while ignoring the impact of physical topology and network latency, resulting in excessive communication overhead for tasks across physical clusters and reduced system performance. By perceiving the physical topology, the system can schedule tasks to computing nodes with lower network latency and closer physical distances based on their needs, thereby reducing performance loss caused by data transmission.

[0147] c) Topology-aware scheduling algorithm

[0148] In order to further optimize resource allocation efficiency, the system introduces a topology-aware scheduling algorithm. This algorithm can intelligently select suitable resource nodes based on factors such as physical topology, network bandwidth between nodes, and latency. Specifically, the scheduling algorithm calculates based on the following key indicators:

[0149] Network delay: The scheduling system monitors the network delay between different nodes and dynamically selects the path with the lowest delay to ensure that the computing task can minimize the waiting time during data transmission.

[0150] Network bandwidth: Different physical clusters may have different network bandwidth limitations. The scheduling algorithm will give priority to nodes with sufficient bandwidth and the ability to carry large amounts of data transmission for task allocation to avoid network congestion.

[0151] Physical distance: The physical distance between nodes directly affects the communication delay. The scheduling system will prioritize assigning tasks to nodes that are closer, thereby reducing scheduling across physical clusters or data centers.

[0152] Through comprehensive evaluation of the above indicators, the topology-aware scheduling algorithm can perform efficient task scheduling within the resource pool or across clusters, avoiding performance bottlenecks caused by disordered scheduling. At the same time, this algorithm is particularly suitable for use in large-scale, high-concurrency computing scenarios, and can greatly improve task execution efficiency.

[0153] d) Reduce the communication overhead of task execution

[0154] In complex computing tasks, especially those involving distributed computing or multi-node parallel computing, data exchange between tasks occurs frequently. If tasks are assigned to nodes that are physically far away and have high communication delays, the data transmission time may far exceed the computing time, resulting in a significant drop in system performance.

[0155] Through the designation of physical service clusters and topology-aware scheduling algorithms, the system can minimize cross-network data transmission during task execution. For tasks that rely on high-frequency data exchange, the scheduling algorithm will centrally distribute these tasks to nodes that are physically adjacent, or even in the same rack or computer room, thereby significantly reducing the additional overhead caused by network communication. This optimization strategy is particularly suitable for the following scenarios:

[0156] High-performance computing (HPC): In high-performance computing tasks such as scientific computing, simulation, and genomics, a large amount of parallel computing and data exchange between nodes are usually required. Topology-aware scheduling can ensure low latency and high bandwidth for data transmission, maximizing computing efficiency.

[0157] Big data processing: In scenarios such as data analysis and stream processing, the amount of data is huge and needs to be transmitted and processed between multiple nodes. Through the optimization of physical service clusters, the system can significantly reduce the transmission time of data between clusters and speed up data processing.

[0158] Through the designation of physical service clusters and topology-aware scheduling, the system can significantly reduce the communication overhead in task execution while ensuring efficient computing, improving overall computing performance, especially in distributed, multi-node large-scale computing scenarios. This method can not only improve the utilization efficiency of computing resources, but also reduce the occupation of network resources, and achieve overall optimization of data center resources.

[0159] In addition, the management system includes a user-friendly interface and visual management functions, as well as localized deployment and data security functions.

[0160] The user-friendly interface and visual management enable administrators and users to efficiently manage and monitor computing resources, which not only improves the intuitiveness of operations, but also greatly enhances the flexibility and scalability of the system by supporting automation and system integration.

[0161] Localized deployment and data security: Localized deployment means deploying the system and all its functional modules, including core computing components and computing models of the intelligent computing center (such as AI reasoning models, large language models, etc.), in the data center within the enterprise or institution. This approach has significant advantages in terms of security, performance, and data privacy, and is particularly suitable for industries and application scenarios that are highly sensitive to data.

[0162] Avoid data leakage: Since the system is deployed in the local data center, all data processing, computing tasks, and AI model reasoning and training are performed locally, and data does not need to be uploaded to an external cloud platform or third-party server. This architecture ensures that data is always in the hands of the user, avoiding the risk of data leakage, and is particularly suitable for organizations that have strict requirements for data security.

[0163] Reduce compliance risks: In many industries, data privacy and security are subject to strict legal and regulatory constraints, such as financial services, healthcare, and government departments. Through local deployment, the system can help enterprises and institutions better comply with relevant regulatory requirements, such as GDPR (General Data Protection Regulation), HIPAA (Health Insurance Portability and Accountability Act), etc., and reduce compliance risks.

[0164] The embodiment of the present invention provides a computing resource management device for an intelligent computing center, see Figure 3 As shown, Figure 3 This is a principle block diagram of a computing resource management device of an intelligent computing center according to an embodiment of the present invention. The computing resource management device 30 includes:

[0165] A prediction module 31 is used to use a large language model for predicting computing power requirements deployed in an intelligent computing center to predict the computing power requirements of the intelligent computing center at a future time;

[0166] A determination module 32 is used to determine a computing power resource scheduling plan from the current moment to the future moment according to the computing power demand;

[0167] The execution module 33 is used to execute the computing resource scheduling plan until the future moment.

[0168] In some embodiments of the present invention, optionally, the determination module 32 is further configured to input the computing power demand into a resource scheduling large language model deployed in an intelligent computing center to determine the computing power resource scheduling solution;

[0169] The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

[0170] In some embodiments of the present invention, optionally, the resource scheduling large language model is trained according to at least one of the following constraints: resource capacity constraint, task requirement constraint, quota constraint, and logical isolation constraint.

[0171] In some embodiments of the present invention, optionally, the prediction module 31 is further configured to receive a scheduling request sent by a user through an interactive terminal, wherein the scheduling request includes user identity information; based on the user identity information, verify whether the user has access rights to access computing resources of an intelligent computing center, and obtain a verification result;

[0172] The prediction module 31 is also used to determine the computing power resource quota of the user in the intelligent computing center if the verification result indicates that the user has the access right, and based on the computing power resource quota, execute the steps of using the computing power demand prediction large language model deployed in the intelligent computing center to predict the computing power demand of the intelligent computing center at a future moment, determine the computing power resource scheduling plan from the current moment to the future moment according to the computing power demand, and execute the computing power resource scheduling plan until the future moment.

[0173] The prediction module 31 is also used for not executing the step of using the large language model for predicting computing power requirements deployed in the intelligent computing center to predict the computing power requirements of the intelligent computing center at a future moment if the verification result indicates that the user does not have the access permission.

[0174] In some embodiments of the present invention, optionally, the computing power resource quota is determined according to a quota constraint condition, and the expression of the quota constraint condition is as follows:

[0175]

[0176] Among them, x u,p,t The allocation of resource type r for user u on project p; The resource quota limit for user u in project p; For all users u.

[0177] The computing resource management device of the intelligent computing center provided in the embodiment of the present application can achieve Figure 1 to Figure 2The various processes implemented by the method embodiment and achieving the same technical effect are not described here to avoid repetition.

[0178] An embodiment of the present invention provides an electronic device 40, see Figure 4 As shown, Figure 4 This is a principle block diagram of an electronic device 40 according to an embodiment of the present invention, including a processor 41, a memory 42, and a program or instruction stored in the memory 42 and executable on the processor 41. When the program or instruction is executed by the processor, the steps in any one of the computing power resource management methods of the intelligent computing center according to the present invention are implemented.

[0179] An embodiment of the present invention provides a readable storage medium, on which programs or instructions are stored. When the programs or instructions are executed by a processor, the various processes of an embodiment of a computing power resource management method for an intelligent computing center such as any of the above-mentioned items are implemented, and the same technical effect can be achieved. To avoid repetition, they will not be repeated here.

[0180] The readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. In some examples, the readable storage medium may be a non-transitory readable storage medium.

[0181] An embodiment of the present invention also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes of the computing power resource management method of the intelligent computing center described above are implemented, and the same technical effect can be achieved. To avoid repetition, they will not be described here.

[0182] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A computing resource management method for an intelligent computing center, characterized in that: include: Step S1: using a large language model for predicting computing power requirements deployed in an intelligent computing center to predict the computing power requirements of the intelligent computing center at a future time; Step S2: Determine a computing resource scheduling plan from the current moment to the future moment according to the computing power demand; Step S3: Execute the computing resource scheduling plan until the future moment.

2. The computing resource management method of the intelligent computing center according to claim 1, characterized in that: The step S2 comprises: Step S21: inputting the computing power demand into a resource scheduling large language model deployed in an intelligent computing center to determine the computing power resource scheduling solution; The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

3. The computing resource management method of the intelligent computing center according to claim 2 is characterized in that: The resource scheduling large language model is trained according to at least one of the following constraints: resource capacity constraint, task requirement constraint, quota constraint, and logical isolation constraint.

4. The computing resource management method of the intelligent computing center according to claim 1, characterized in that: The step S1 includes: Step S01: receiving a scheduling request sent by a user through an interactive terminal, wherein the scheduling request includes user identity information; based on the user identity information, verifying whether the user has access rights to access computing resources of an intelligent computing center, and obtaining a verification result; Step S02: if the verification result indicates that the user has the access right, determine the computing resource quota of the user in the intelligent computing center, and execute the steps S1, S2 and S3 based on the computing resources of the computing resource quota; Step S03: If the verification result indicates that the user does not have the access authority, step S1 is not executed.

5. The computing resource management method of the intelligent computing center according to claim 4 is characterized in that: The computing power resource quota is determined according to the quota constraint condition, and the expression of the quota constraint condition is as follows: Among them, x u,p,r The allocation of resource type r for user u on project p; The resource quota limit for user u in project p; For all users u.

6. A computing resource management device for an intelligent computing center, characterized in that: include: A prediction module, used to use a large language model for predicting computing power requirements deployed in an intelligent computing center to predict the computing power requirements of the intelligent computing center at future moments; A determination module, used to determine a computing power resource scheduling plan from the current moment to the future moment according to the computing power demand; An execution module is used to execute the computing resource scheduling plan until the future moment.

7. The computing resource management device of the intelligent computing center according to claim 6, characterized in that: The determination module is further used to input the computing power demand into a resource scheduling large language model deployed in the intelligent computing center to determine the computing power resource scheduling plan; The resource scheduling large language model is trained according to a first objective function, and the first objective function is determined according to resource usage cost and resource usage energy consumption.

8. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps in the computing power resource management method of the intelligent computing center as described in any one of claims 1 to 5.

9. A readable storage medium, characterized in that: The readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the steps in the computing resource management method of the intelligent computing center as described in any one of claims 1 to 5 are implemented.

10. A computer program product, characterized in that It comprises computer instructions, which, when executed by a processor, implement the steps of the computing resource management method of the intelligent computing center as described in any one of claims 1 to 5.

Images