Intelligent resource allocation optimization method and system for electric power hyper-converged computing platform

Abstract

The invention relates to an intelligent resource allocation optimization method for an electric power hyper-converged computing platform. The method comprises the following steps: S1, obtaining operation data of the electric power hyper-converged computing platform; s2, constructing a load prediction model based on LSTM, and performing load prediction according to the collected operation data; s3, setting an optimization target for a resource allocation problem of the hyper-fusion platform, and constructing an optimization target function; s4, according to the predicted load data and based on the objective function, introducing an intelligent algorithm to optimize resource allocation, and obtaining an optimized resource allocation scheme; and S5, according to the optimized resource allocation scheme, in combination with a virtualization technology and a container arrangement technology, mapping the optimized resource allocation scheme into an actual operation environment. According to the invention, the resource utilization rate can be effectively improved in the power hyper-fusion platform, the system energy consumption is reduced, and the task timeliness demand is ensured.

Description

Technical Field

[0001] The present invention relates to the field of resource allocation, and in particular to an intelligent resource allocation optimization method and system for an electric power hyper-converged computing platform. Background Art

[0002] With the digital and intelligent transformation of the power industry, the power system has put forward higher requirements for the performance and resource management of computing platforms. The power hyper-converged computing platform is an efficient computing platform that integrates computing, storage, network and other resources into a unified architecture, which can support key businesses such as power dispatching, big data analysis, and real-time prediction. However, due to the dynamic and time-varying nature of power system workloads (such as real-time dispatching requirements during peak power consumption periods and data analysis tasks during low power consumption periods), traditional static resource allocation and fixed strategies are stretched when dealing with these complex scenarios, often leading to problems such as low resource utilization, delayed task execution, and excessive energy consumption. To this end, combining artificial intelligence technology to predict power loads and using intelligent optimization algorithms for dynamic resource scheduling have become the key to solving resource management bottlenecks. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide an intelligent resource allocation optimization method and system for a power hyper-converged computing platform, which can effectively improve resource utilization, reduce system energy consumption, and ensure task timeliness requirements in the power hyper-converged computing platform.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] An intelligent resource allocation optimization method for an electric power hyper-converged computing platform comprises the following steps:

[0006] S1: Obtaining the operation data of the power hyper-converged computing platform;

[0007] S2: Build a load prediction model based on LSTM and perform load prediction based on the collected operation data;

[0008] S3: Set optimization goals for the resource allocation problem of the hyper-converged platform and construct optimization objective functions;

[0009] S4: According to the predicted load data and based on the above objective function, an intelligent algorithm is introduced to optimize resource allocation to obtain an optimized resource allocation plan;

[0010] S5: Based on the optimized resource allocation plan, combined with virtualization technology and container orchestration technology, the optimized resource allocation plan is mapped to the actual operating environment.

[0011] Furthermore, a load prediction model is built based on LSTM, and load prediction is performed based on the collected operation data, as follows:

[0012] Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t ,BW t}、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t} and user demand data: D USR = {Peak t ,Priority t}, and obtain the corresponding load data Y;

[0013] Among them, CPU t is the CPU utilization; GPU t is GPU utilization; Mem t is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput;

[0014] QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module; DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t It is the peak and valley period label submitted by the user task; Priorityt is the distribution ratio of task priority;

[0015] Construct a regular time series X from historical data according to time step t:

[0016] X={x t-n ,x t-n+1 ,…,x t};

[0017] Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns;

[0018] Divide the time series X and target Y data into training set and validation set in proportion;

[0019] A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

[0020] Furthermore, the objective function is optimized as follows:

[0021] According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay:

[0022] maximize F=α1U avg -α2P total -α3T delay

[0023] Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively;

[0024] Resource allocation constraints include the following:

[0025] Total resource constraints: The resource allocation of each node must not exceed its available resources;

[0026] Task allocation constraint: Each task must be assigned to and only to one node;

[0027] Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay;

[0028] Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

[0029] Furthermore, the intelligent algorithm adopts a genetic algorithm, as follows:

[0030] Randomly generate P size chromosomes, as a population;

[0031] Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints;

[0032] The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function;

[0033] Select excellent chromosomes to enter the next generation based on fitness values;

[0034] Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring;

[0035] Set the crossover probability p c , exchange tasks in proportion;

[0036] Randomly modify the assigned nodes of a task;

[0037] Set the mutation probability p m , to avoid premature population maturation and falling into local optimum;

[0038] When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

[0039] Furthermore, by combining virtualization technology with container orchestration technology, the optimized resource allocation scheme is mapped to the actual operating environment, as follows:

[0040] According to the optimized resource allocation plan, the resource requirements of each task are decomposed into the following core indicators: Requested resources: the minimum operating resources of the task; Maximum limit: the maximum resources that the task can occupy; Priority: the importance and preemptibility of the task;

[0041] Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity.

[0042] Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with abundant resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node or existing tasks are migrated to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded during runtime.

[0043] An intelligent resource allocation optimization system for an electric power hyper-converged computing platform comprises a data acquisition module, a load prediction module, a resource allocation module and a virtualization and containerization scheduling module; the data acquisition module is used to obtain the operating data of the electric power hyper-converged computing platform; the load prediction module constructs a load prediction model based on LSTM, and performs load prediction according to the collected operating data; the resource allocation module sets an optimization goal for the resource allocation problem of the hyper-converged platform, constructs an optimization objective function, and introduces an intelligent algorithm to optimize resource allocation based on the predicted load data and the above objective function to obtain an optimized resource allocation plan; the virtualization and containerization scheduling module maps the optimized resource allocation plan to the actual operating environment according to the optimized resource allocation plan in combination with virtualization technology and container orchestration technology.

[0044] Furthermore, a load prediction model is built based on LSTM, and load prediction is performed based on the collected operation data, as follows:

[0045] Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t ,BW t}、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t} and user demand data: D USR = {Peak t ,Priority t}, and obtain the corresponding load data Y;

[0046] Among them, CPU t is the CPU utilization; GPU t is GPU utilization; Mem t is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput; QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module; DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t It is the peak and valley period label submitted by the user task; Priorityt is the distribution ratio of task priority;

[0047] Construct a regular time series X from historical data according to time step t:

[0048] X={x t-n ,x t-n+1 ,…,x t};

[0049] Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns;

[0050] Divide the time series X and target Y data into training set and validation set in proportion;

[0051] A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

[0052] Furthermore, the objective function is optimized as follows:

[0053] According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay:

[0054] maximize F=α1U avg -α2P total -α3T delay

[0055] Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively;

[0056] Resource allocation constraints include the following:

[0057] Total resource constraints: The resource allocation of each node must not exceed its available resources;

[0058] Task allocation constraint: Each task must be assigned to and only to one node;

[0059] Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay;

[0060] Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

[0061] Furthermore, the intelligent algorithm adopts a genetic algorithm, as follows:

[0062] Randomly generate P size chromosomes, as a population;

[0063] Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints;

[0064] The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function;

[0065] Select excellent chromosomes to enter the next generation based on fitness values;

[0066] Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring;

[0067] Set the crossover probability p c , exchange tasks in proportion;

[0068] Randomly modify the assigned nodes of a task;

[0069] Set the mutation probability p m , to avoid premature population maturation and falling into local optimum;

[0070] When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

[0071] Furthermore, by combining virtualization technology with container orchestration technology, the optimized resource allocation scheme is mapped to the actual operating environment, as follows:

[0072] According to the optimized resource allocation plan, the resource requirements of each task are decomposed into the following core indicators: Requested resources: the minimum operating resources of the task; Maximum limit: the maximum resources that the task can occupy; Priority: the importance and preemptibility of the task;

[0073] Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity.

[0074] Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with sufficient resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node or existing tasks are migrated to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded at runtime:

[0075] 1. The present invention can effectively improve resource utilization, reduce system energy consumption, and ensure task timeliness requirements in the power hyper-convergence platform;

[0076] 2. The present invention constructs an optimization objective function based on the balance of multiple objectives such as resource utilization, delay and energy consumption, optimizes resource allocation by combining heuristic algorithms, mathematical programming and deep learning technology, effectively improves the operating efficiency and task responsiveness of the hyper-converged platform, and transforms the resource allocation problem into an optimization problem by introducing a genetic algorithm. The resource allocation scheme is dynamically adjusted by using load prediction data and multi-objective optimization objective functions, effectively improving the resource utilization of the hyper-converged platform while meeting multiple constraints such as delay, energy consumption and priority.

[0077] 3. Based on the optimized resource allocation scheme, the present invention can combine virtualization technology and container orchestration technology Kubernetes to build a distributed resource scheduling system, and complete efficient resource allocation through dynamic deployment, task migration and load balancing, thereby achieving improved resource utilization, flexible dynamic expansion and contraction, and effective control of task delays. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a flow chart of the method of the present invention;

[0079] Figure 2 FIG. 4 is a system architecture diagram in one embodiment of the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0081] refer to Figure 1 In this embodiment, a method for optimizing intelligent resource allocation of a power hyper-converged computing platform includes the following steps:

[0082] S1: Obtaining the operation data of the power hyper-converged computing platform;

[0083] S2: Build a load prediction model based on LSTM and perform load prediction based on the collected operation data;

[0084] S3: Set optimization goals for the resource allocation problem of the hyper-converged platform and construct optimization objective functions;

[0085] S4: According to the predicted load data and based on the above objective function, an intelligent algorithm is introduced to optimize resource allocation to obtain an optimized resource allocation plan;

[0086] S5: Based on the optimized resource allocation plan, combined with virtualization technology and container orchestration technology, the optimized resource allocation plan is mapped to the actual operating environment.

[0087] In this embodiment, a load prediction model is constructed based on LSTM, and load prediction is performed according to the collected operation data, as follows:

[0088] Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t ,BW t}、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t} and user demand data: D USR = {Peak t ,Priority t}, and obtain the corresponding load data Y;

[0089] Among them, CPU t is the CPU utilization; GPU t is the GPU utilization (load rate, video memory occupancy, etc.); Memt is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput; QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module (such as CPU / GPU usage); DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t The peak and valley period labels submitted by the user tasks (e.g., peak period is 1, valley period is 0); Priorityt is the distribution ratio of task priority;

[0090] Construct a regular time series X from historical data according to time step t:

[0091] X={x t-n ,x t-n+1 ,…,x t};

[0092] Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns;

[0093] Divide the time series X and target Y data into training set and validation set in proportion;

[0094] A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

[0095] In this embodiment, the objective function is optimized as follows:

[0096] According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay:

[0097] maximize F=α1U avg -α2P total -α3T delay

[0098] Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively;

[0099] Resource allocation constraints include the following:

[0100] Total resource constraints: The resource allocation of each node must not exceed its available resources;

[0101] Task allocation constraint: Each task must be assigned to and only to one node;

[0102] Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay;

[0103] Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

[0104] In this embodiment, the intelligent algorithm adopts a genetic algorithm, which is as follows:

[0105] Randomly generate P size chromosomes (initial solution), as the population;

[0106] Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints;

[0107] The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function;

[0108] Select excellent chromosomes to enter the next generation based on fitness values ​​(such as roulette selection or tournament selection);

[0109] Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring;

[0110] Set the crossover probability p c , exchange tasks in proportion;

[0111] Randomly modify the assigned nodes of a task;

[0112] Set the mutation probability p m , to avoid premature population maturation and falling into local optimum;

[0113] When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

[0114] In this embodiment, virtualization technology and container orchestration technology are combined to map the optimized resource allocation scheme to the actual operating environment, as follows:

[0115] According to the optimized resource allocation plan, the resource requirements (CPU, memory, GPU, etc.) of each task are decomposed into the following core indicators: Requests: the minimum running resources of the task (for example, CPU: 2 cores, memory: 4GB); Limits: the maximum resources that a task can occupy (for example, CPU: 6 cores, memory: 12GB); Priority: the importance and preemptiveness of the task (for example, high-priority tasks will preempt node resources first);

[0116] Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity.

[0117] Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with abundant resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node (expanding virtual machines or container capacity) or migrating existing tasks to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded during runtime. For example, vCPU or memory is added to the task. In a virtualized environment, the configuration of the virtual machine can be adjusted in real time, such as adjusting memory and CPU usage; horizontal expansion (adding task replicas).

[0118] refer to Figure 2 In this embodiment, an intelligent resource allocation optimization system for an electric power hyper-converged computing platform is also provided, including a data acquisition module, a load prediction module, a resource allocation module and a virtualization and containerization scheduling module; the data acquisition module is used to obtain the operating data of the electric power hyper-converged computing platform; the load prediction module constructs a load prediction model based on LSTM, and performs load prediction based on the collected operating data; the resource allocation module sets an optimization goal for the resource allocation problem of the hyper-converged platform, constructs an optimization objective function, and introduces an intelligent algorithm to optimize the resource allocation based on the predicted load data and the above objective function to obtain an optimized resource allocation plan; the virtualization and containerization scheduling module maps the optimized resource allocation plan to the actual operating environment based on the optimized resource allocation plan and combines virtualization technology with container orchestration technology.

[0119] In this embodiment, a load prediction model is constructed based on LSTM, and load prediction is performed according to the collected operation data, as follows:

[0120] Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t,BW t}、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t} and user demand data: D USR = {Peak t ,Priority t}, and obtain the corresponding load data Y;

[0121] Among them, CPU t is the CPU utilization; GPU t is the GPU utilization (load rate, video memory occupancy, etc.); Mem t is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput; QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module (such as CPU / GPU usage); DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t The peak and valley period labels submitted by the user tasks (e.g., peak period is 1, valley period is 0); Priorityt is the distribution ratio of task priority;

[0122] Construct a regular time series X from historical data according to time step t:

[0123] X={x t-n ,x t-n+1 ,…,x t};

[0124] Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns;

[0125] Divide the time series X and target Y data into training set and validation set in proportion;

[0126] A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

[0127] In this embodiment, the objective function is optimized as follows:

[0128] According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay:

[0129] maximize F=α1U avg -α2P total -α3T delay

[0130] Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively;

[0131] Resource allocation constraints include the following:

[0132] Total resource constraints: The resource allocation of each node must not exceed its available resources;

[0133] Task allocation constraint: Each task must be assigned to and only to one node;

[0134] Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay;

[0135] Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

[0136] In this embodiment, the intelligent algorithm adopts a genetic algorithm, which is as follows:

[0137] Randomly generate P size chromosomes (initial solution), as the population;

[0138] Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints;

[0139] The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function;

[0140] Select excellent chromosomes to enter the next generation based on fitness values ​​(such as roulette selection or tournament selection);

[0141] Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring;

[0142] Set the crossover probability p c , exchange tasks in proportion;

[0143] Randomly modify the assigned nodes of a task;

[0144] Set the mutation probability p m, to avoid premature population maturation and falling into local optimum;

[0145] When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

[0146] In this embodiment, virtualization technology and container orchestration technology are combined to map the optimized resource allocation scheme to the actual operating environment, as follows:

[0147] According to the optimized resource allocation plan, the resource requirements (CPU, memory, GPU, etc.) of each task are decomposed into the following core indicators: Requests: the minimum running resources of the task (for example, CPU: 2 cores, memory: 4GB); Limits: the maximum resources that a task can occupy (for example, CPU: 6 cores, memory: 12GB); Priority: the importance and preemptiveness of the task (for example, high-priority tasks will preempt node resources first);

[0148] Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity.

[0149] Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with abundant resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node (expanding virtual machines or container capacity) or migrating existing tasks to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded during runtime. For example, vCPU or memory is added to the task. In a virtualized environment, the configuration of the virtual machine can be adjusted in real time, such as adjusting memory and CPU usage; horizontal expansion (adding task replicas).

[0150] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0152] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0154] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. An intelligent resource allocation optimization method for a power hyper-converged computing platform, characterized in that: The following steps are involved: S1: Obtaining the operation data of the power hyper-converged computing platform; S2: Build a load prediction model based on LSTM and perform load prediction based on the collected operation data; S3: Set optimization goals for the resource allocation problem of the hyper-converged platform and construct optimization objective functions; S4: According to the predicted load data and based on the above objective function, an intelligent algorithm is introduced to optimize resource allocation to obtain an optimized resource allocation plan; S5: Based on the optimized resource allocation plan, combined with virtualization technology and container orchestration technology, the optimized resource allocation plan is mapped to the actual operating environment.

2. The intelligent resource allocation optimization method for a power hyper-converged computing platform according to claim 1 is characterized in that: The load prediction model is constructed based on LSTM, and the load prediction is performed according to the collected operation data, as follows: Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t ,BW t }、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t } and user demand data: D USR = {Peak t ,Priority t }, and obtain the corresponding load data Y; Among them, CPU t is CPU utilization; GPU t is GPU utilization; Mem t is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput; QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module; DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t It is the peak and valley period label submitted by the user task; Priorityt is the distribution ratio of task priority; Construct a regular time series X from historical data according to time step t: X={x t-n ,x t-n+1 ,…,x t }; Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns; Divide the time series X and target Y data into training set and validation set in proportion; A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

3. The intelligent resource allocation optimization method for a power hyper-converged computing platform according to claim 1 is characterized in that: The optimization objective function is as follows: According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay: maximize F=α1U aυg -α2P total -α3T delay Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively; Resource allocation constraints include the following: Total resource constraints: The resource allocation of each node must not exceed its available resources; Task allocation constraint: Each task must be assigned to and only to one node; Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay; Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

4. The intelligent resource allocation optimization method for a power hyper-converged computing platform according to claim 1 is characterized in that: The intelligent algorithm adopts a genetic algorithm, which is as follows: Randomly generate P size chromosomes, as a population; Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints; The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function; Select excellent chromosomes to enter the next generation based on fitness values; Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring; Set the crossover probability pc and swap tasks in proportion; Randomly modify the assigned nodes of a task; Set the mutation probability pm to avoid premature population maturation and falling into local optimum; When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

5. The intelligent resource allocation optimization method for a power hyper-converged computing platform according to claim 1 is characterized in that: The virtualization technology and container orchestration technology are combined to map the optimized resource allocation scheme to the actual operating environment, as follows: According to the optimized resource allocation plan, the resource requirements of each task are decomposed into the following core indicators: Requested resources: the minimum operating resources of the task; Maximum limit: the maximum resources that the task can occupy; Priority: the importance and preemptibility of the task; Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity. Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with abundant resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node or existing tasks are migrated to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded during runtime.

6. An intelligent resource allocation optimization system for a power hyper-converged computing platform, characterized in that: It includes a data acquisition module, a load prediction module, a resource allocation module and a virtualization and containerization scheduling module; the data acquisition module is used to obtain the operating data of the power hyper-converged computing platform; the load prediction module builds a load prediction model based on LSTM, and performs load prediction based on the collected operating data; the resource allocation module sets optimization goals for the resource allocation problem of the hyper-converged platform, constructs an optimization objective function, and introduces an intelligent algorithm to optimize resource allocation based on the predicted load data and the above objective function to obtain an optimized resource allocation plan; the virtualization and containerization scheduling module maps the optimized resource allocation plan to the actual operating environment based on the optimized resource allocation plan and combines virtualization technology with container orchestration technology.

7. The intelligent resource allocation optimization system for a power hyper-converged computing platform according to claim 6, characterized in that: The load prediction model is constructed based on LSTM, and the load prediction is performed according to the collected operation data, as follows: Get historical data including hardware resource data D HW ={CPU t ,GPU t, Mem t ,IO t ,BW t }、Business load data D BUS = {QueueLength t ,HPC t ,DataAnalysis t } and user demand data: D USR = {Peak t ,Priority t }, and obtain the corresponding load data Y; Among them, CPU t is CPU utilization; GPU t is GPU utilization; Mem t is the memory usage; IO t BW is the disk I / O read and write rate; t is the network bandwidth throughput; QueueLength t is the task queue length; HPC t is the task load of the high-performance computing module; DataAnalysis is the input / processing data volume of the task in the data analysis module; Peak t It is the peak and valley period label submitted by the user task; Priorityt is the distribution ratio of task priority; Construct a regular time series X from historical data according to time step t: X={x t-n ,x t-n+1 ,…,x t }; Among them, x t =[D HW (t),D BUS (t),D USR (t)] is the feature vector at the current moment, which includes hardware data, task load data, and user behavior patterns; Divide the time series X and target Y data into training set and validation set in proportion; A load prediction model is built based on LSTM, trained based on the training set, and optimized using the optimization algorithm Adam to obtain the trained load prediction model.

8. The intelligent resource allocation optimization system for a power hyper-converged computing platform according to claim 6, characterized in that: The optimization objective function is as follows: According to the goal setting, the multi-objective optimization function is constructed by comprehensively considering resource utilization, energy consumption and task delay: maximize F=α1U aυg -α2P total -α3T delay Among them, U avg is the average resource utilization of the node; P total is the total energy consumption; T delay is the total task delay; α1, α2, α3 are the weight coefficients of resource utilization, energy consumption, and delay respectively; Resource allocation constraints include the following: Total resource constraints: The resource allocation of each node must not exceed its available resources; Task allocation constraint: Each task must be assigned to and only to one node; Delay satisfies the constraint: the total delay of the task must not exceed its required limit delay; Priority guarantee constraint: high-priority tasks are allocated sufficient resources first, and low-priority tasks must not affect high-priority tasks.

9. The intelligent resource allocation optimization system for a power hyper-converged computing platform according to claim 6, characterized in that: The intelligent algorithm adopts a genetic algorithm, which is as follows: Randomly generate P size chromosomes, as a population; Each chromosome corresponds to a resource allocation scheme that satisfies resource allocation constraints; The optimization goal of the genetic algorithm is to maximize the fitness function F(x), which converts the previous objective function into a fitness function; Select excellent chromosomes to enter the next generation based on fitness values; Select two chromosomes X1 and X2 from the parent generation, exchange some task allocation schemes to generate offspring; Set the crossover probability p c , exchange tasks in proportion; Randomly modify the assigned nodes of a task; Set the mutation probability p m , to avoid premature population maturation and falling into local optimum; When the maximum number of iterations is reached or the population fitness value converges, the algorithm is terminated and the optimized resource allocation plan is obtained.

10. The intelligent resource allocation optimization system for a power hyper-converged computing platform according to claim 6, characterized in that: The virtualization technology and container orchestration technology are combined to map the optimized resource allocation scheme to the actual operating environment, as follows: According to the optimized resource allocation plan, the resource requirements of each task are decomposed into the following core indicators: Requested resources: the minimum operating resources of the task; Maximum limit: the maximum resources that the task can occupy; Priority: the importance and preemptibility of the task; Based on the resource allocation results of the optimization plan, tasks are bound to specific compute nodes. On Kubernetes, tasks are fixed to target nodes by using NodeSelector or node affinity. Based on the remaining resources of each node, high-priority tasks are scheduled to nodes with abundant resources first, while meeting the actual needs of the tasks. If the node resources are insufficient, additional resources are allocated to the node or existing tasks are migrated to meet the task operation. If the container output prompt of a task exceeds the load demand, its resources are expanded during runtime.

Images