A heterogeneous computing resource pooling management system and method

Abstract

The application relates to the technical field of resource pooling, and discloses a heterogeneous computing resource pooling management system and method, which comprises the following steps: according to real-time state data of heterogeneous resources, the state of each computing resource in a resource pool is coded to obtain resource coding; in response to a computing task, a communication dependency graph is constructed by analyzing the communication between computing resources, and a first resource combination and a second resource combination are screened out; in the process of the computing task, the first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resources, and the first resource pool and the second resource pool are dynamically adjusted; after the computing task is completed, the first resource combination and the second resource combination are decoupled and put into the corresponding resource pool; the application screens and dynamically couples the computing resources, so that the heterogeneous computing resources can be more fully utilized, resource waste is reduced, and efficient pooling management of the heterogeneous computing resources is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource pooling, more particularly to a heterogeneous computing resource pooling management system and method. BACKGROUND

[0002] At present, heterogeneous computing resources play a key role in various computing scenarios, including different types of computing units such as CPU, GPU, FPGA, etc., and are widely used in cloud computing, big data analysis, artificial intelligence training, and high-performance computing fields. In the development process of information technology, the scale and complexity of computing resources are required to be higher, and the existing resource pooling management method cannot adapt to the dynamic combination and adjustment of computing resources for different computing tasks, resulting in low computing performance and resource utilization.

[0003] The existing technology has the following problems: in the process of computing resource screening, selection is made based on a single resource attribute, resulting in low communication efficiency of the screened resource combination and affecting the processing speed of the computing task; in the process of computing task, the screening and combination of heterogeneous computing resources are performed through a fixed combination mechanism, which is difficult to adapt to the dynamic change requirements of the task; directly using the screened resources to perform the corresponding computing task lacks dynamic combination and adjustment of computing resources, resulting in low resource utilization; to solve at least one of the above problems, the present application proposes a heterogeneous computing resource pooling management system and method. SUMMARY

[0004] In view of the deficiencies of the prior art, the purpose of the present application is to provide a heterogeneous computing resource pooling management system and method, which can effectively solve the problems in the background art. The specific technical solution of the present application is as follows:

[0005] A heterogeneous computing resource pooling management method, comprising:

[0006] According to the real-time state data of the heterogeneous resources, the state of each computing resource in the resource pool is encoded to obtain a resource code, wherein the resource pool includes a first resource pool storing resource codes greater than a preset encoding threshold and a second resource pool storing resource codes less than or equal to the preset encoding threshold;

[0007] In response to a computing task, a communication dependency graph is constructed by analyzing the communication between computing resources through a preset resource combination model, a first resource combination is screened from the first resource pool, and a second resource combination is screened from the second resource pool;

[0008] In the process of computing task, the first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resources, and the utilization rate of the computing resources is analyzed to dynamically adjust the first resource pool and the second resource pool;

[0009] After the computing task ends, the first resource combination and the second resource combination are decoupled, and according to the state of the decoupled computing resources, the corresponding resource pool is put into, so as to pool management of the heterogeneous computing resources.

[0010] Specifically, the state of each computing resource in the resource pool is encoded according to the real-time state data of the heterogeneous resources, and resource encoding is obtained, including:

[0011] The real-time state data of the heterogeneous resources is state feature extracted, and a state vector is constructed.

[0012] The state vector is mapped to the pre-constructed resource double chain structure for encoding, and the resource encoding of each computing resource is obtained.

[0013] The computing resource with resource encoding greater than the preset encoding threshold is put into the first resource pool, and the computing resource with resource encoding less than or equal to the preset encoding threshold is put into the second resource pool.

[0014] Specifically, the state vector is mapped to the pre-constructed resource double chain structure for encoding, and the resource encoding of each computing resource is obtained, including:

[0015] According to the dimension of the state vector, a feature space corresponding to the dimension is constructed, the state vector is mapped to the feature space, and a feature sequence is generated.

[0016] The resource double chain structure is constructed by combining the historical state data of the computing resources and the feature sequence, and the resource double chain structure includes a first resource chain constructed according to the feature sequence and a second resource chain constructed by complementary conversion of the feature sequence.

[0017] Based on the real-time resource load of the resource pool, the information entropy value of each computing resource at the corresponding position in the resource double chain structure is calculated, and the corresponding resource encoding is obtained.

[0018] Specifically, in response to the computing task, a communication dependency graph is constructed by analyzing the communication between the computing resources through a preset resource combination model, a first resource combination is selected from the first resource pool, and a second resource combination is selected from the second resource pool, including:

[0019] In response to the computing task, the communication dependency value between the computing resources is analyzed by combining the resource encoding of the computing resources, and a communication dependency graph and a task communication constraint are generated.

[0020] Through the preset resource combination model, resource nodes satisfying the task communication constraint are selected from the communication dependency graph corresponding to the first resource pool, and the first resource combination is obtained, and resource nodes connected with the resource nodes in the first resource combination are selected from the communication dependency graph corresponding to the second resource pool, and the second resource combination is obtained.

[0021] According to the screened first resource combination and the second resource combination, the communication dependency graph is updated in real time.

[0022] Specifically, the communication dependency value between the computing resources is analyzed in response to the computing task in combination with the resource encoding of the computing resources, and a communication dependency graph and a task communication constraint are generated, including:

[0023] According to the resource encoding of the computing resources and the topology structure between the resource nodes, a first communication value between the computing resources is calculated, and the first communication value is corrected by a stability coefficient in the resource encoding to obtain a communication dependency value;

[0024] An association edge is established between the resource nodes with the communication dependency value greater than a preset dependency threshold, and a communication dependency graph is constructed;

[0025] In response to the computing task, the dependency between the resources corresponding to the task is analyzed, and a task communication constraint is generated in combination with the communication mode corresponding to the task type.

[0026] Specifically, during the process of the computing task, the first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resources, and the utilization rate of the computing resources is analyzed to dynamically adjust the first resource pool and the second resource pool, including:

[0027] During the process of the computing task, a task performance vector is constructed by analyzing the real-time performance indicators of the task;

[0028] The computing resources are dynamically coupled in combination with the task performance vector and the communication dependency values of the computing resources in the first resource combination and the second resource combination to obtain coupled resources;

[0029] The utilization rate of the coupled resources and the resource encoding difference of the computing resources are analyzed to dynamically adjust the first resource pool and the second resource pool.

[0030] Specifically, the computing resources are dynamically coupled in combination with the task performance vector and the communication dependency values of the computing resources in the first resource combination and the second resource combination to obtain coupled resources, including:

[0031] The coupling weight between the corresponding computing resources is calculated in combination with the task performance vector, the real-time state of the computing resources, and the communication dependency values of the computing resources in the first resource combination and the second resource combination;

[0032] The computing resources with the coupling weight greater than a preset coupling threshold are screened out in the first resource combination to obtain first coupled resources;

[0033] screening out the computing resources with the coupling weight greater than the preset coupling threshold from the first coupling resource in the second resource combination, to obtain a second coupling resource;

[0034] coupling the first coupling resource and the second coupling resource, to obtain a coupling resource.

[0035] Specifically, the utilization rate of the coupling resource and the resource code difference of the computing resource are analyzed, and the first resource pool and the second resource pool are dynamically adjusted, including:

[0036] The utilization rate of the coupling resource is analyzed, and if the utilization rate of the first coupling resource in the coupling resource is greater than a preset utilization rate threshold, an extended resource with the same structure as the computing resource in the first coupling resource and the highest resource code is selected from the second resource pool.

[0037] The extended resource is put into the first resource pool, the first resource pool is expanded, and the second resource pool is updated;

[0038] If the absolute value of the continuous resource code difference in the second resource pool is greater than a preset code difference threshold, the computing resource with the lower resource code in the continuous computing resource is put into the cooling resource area in the second resource pool.

[0039] Specifically, after the computing task is completed, the first resource combination and the second resource combination are decoupled, and according to the state of the decoupled computing resource, the corresponding resource pool is put into, to pool the heterogeneous computing resource, including:

[0040] After the computing task is completed, the state of the computing resource in the first resource combination and the second resource combination is analyzed, and the corresponding decoupling degree is calculated;

[0041] According to the decoupling degree from high to low, the computing resource is decoupled to obtain a decoupled resource;

[0042] The state of the decoupled resource is analyzed and calculated to obtain the corresponding resource code, and the resource code is put into the corresponding resource pool.

[0043] A heterogeneous computing resource pooling management system for realizing the heterogeneous computing resource pooling management method, including:

[0044] A resource code module encodes the state of each computing resource in the resource pool according to the real-time state data of the heterogeneous resource, to obtain a resource code;

[0045] A resource screening module screens out a first resource combination from the first resource pool and a second resource combination from the second resource pool in response to a computing task by analyzing the communication between the computing resources through a preset resource combination model to construct a communication dependency graph;

[0046] The resource coupling module dynamically couples the first resource combination and the second resource combination according to the real-time state of the computing resources during the computing task, and analyzes the utilization rate of the computing resources to dynamically adjust the first resource pool and the second resource pool.

[0047] The resource decoupling module decouples the first resource combination and the second resource combination after the computing task is completed, and puts the decoupled computing resources into the corresponding resource pool according to the state of the decoupled computing resources, so as to pool manage the heterogeneous computing resources.

[0048] The application has the following beneficial effects: the state vector of the heterogeneous resources is mapped into the resource double-chain structure for coding, the resources are respectively put into different resource pools according to the resource coding, the corresponding resource combination is respectively screened from the resource pools according to the computing task and the communication condition between the computing resources, the resources in the resource combination are dynamically coupled during the task computing, and the utilization rate of the resources and the resource coding difference are analyzed to expand and adjust the resource pool, the computing resources are decoupled and re-poured into the corresponding resource pool after the computing task is completed; the computing resources are screened and dynamically coupled by the resource coding combined with the communication condition between the computing resources, which can quickly respond to the computing task of the user, so that the heterogeneous computing resources can be more fully utilized, the resource waste is reduced, and the efficient pool management of the heterogeneous computing resources is realized. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The work flow chart of the heterogeneous computing resource pool management method in the embodiment of the application is shown in the figure.

[0050] Figure 2 The schematic diagram of the communication dependency graph in the embodiment of the application is shown in the figure.

[0051] Figure 3 The schematic diagram of the coupling resource screening process in the embodiment of the application is shown in the figure.

[0052] Figure 4 The schematic diagram of the resource pool dynamic adjustment process in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0053] The application will be further described in detail below with reference to the drawings and embodiments. In the embodiments of the application, the words such as "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the application should not be interpreted as being more preferred or having more advantages than other embodiments or design schemes. In fact, the words such as "exemplary" or "for example" are intended to present the relevant concept in a specific manner.

[0054] Hereinafter, the terms "first", "second", and the like are generic terms and are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0055] Reference Figure 1 As shown in the specific embodiment of the heterogeneous computing resource pooling management method of the present application, it comprises:

[0056] S101, according to the real-time state data of the heterogeneous resources, the state of each computing resource in the resource pool is coded to obtain the resource code, wherein the resource pool includes a first resource pool storing the resource code greater than the preset coding threshold and a second resource pool storing the resource code less than or equal to the preset coding threshold;

[0057] S102, in response to the computing task, the communication situation between the computing resources is analyzed by the preset resource combination model to construct a communication dependency graph, the first resource combination is selected from the first resource pool, and the second resource combination is selected from the second resource pool;

[0058] S103, in the process of the computing task, the first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resources, and the utilization rate of the computing resources is analyzed to dynamically adjust the first resource pool and the second resource pool;

[0059] S104, after the computing task is completed, the first resource combination and the second resource combination are decoupled, and according to the state of the decoupled computing resources, they are put into the corresponding resource pool to perform the pooling management of the heterogeneous computing resources.

[0060] In the pooling management of heterogeneous computing resources, the real-time states of different structures of computing resources are collected, and the resources are divided into a first resource pool and a second resource pool according to the state vectors of the resources and the coding results. The communication dependency graph is constructed by analyzing the communication between the computing resources, the corresponding resource combinations are selected from the two resource pools by analyzing the computing tasks, and the resource combinations are dynamically coupled during the task execution process. The resource pool is expanded and adjusted according to the resource utilization, and the resource combinations are decoupled after the task is completed. The resource state after decoupling is re-evaluated and returned to the corresponding resource pool according to the resource state. Through the classification management and screening of resources, the appropriate resource combination can be quickly matched for the computing task, the response speed of the task is improved, the dynamic coupling of resources can optimize the resource cooperation mode, improve the processing efficiency and quality of the task, and the dynamic adjustment of the resource pool can make full use of resources in different states, thereby improving the utilization rate of heterogeneous computing resources.

[0061] In the embodiment, the heterogeneous computing resources have different performance, architecture and real-time state, including but not limited to the processing speed of CPU, the occupancy rate of memory, and the computing power of GPU. The computing resources are coded according to the real-time state data of each computing resource, the state information is converted into quantized resource code, the state of the computing resources is analyzed, and the resources are divided into different resource pools based on the resource state. The resources can be screened and managed according to the demand of the computing task. By coding and classifying the resource state, the performance state of the resource can be quickly distinguished to provide an accurate state reference for resource screening. The resources are divided into different resource pools, and the heterogeneous computing resources in different states can be managed and scheduled quickly.

[0062] The embodiment considers that different computing tasks have different demands for the performance and communication ability of computing resources. When a computing task arrives, the communication between the computing resources is analyzed by a preset resource combination model to construct a communication dependency graph, which reflects the communication relationship and communication cost between the computing resources. Based on the communication dependency graph, the resource combinations that meet the demand of the computing task are screened from the resource pool to ensure efficient execution of the task. By constructing the communication dependency graph, the communication relationship between different resources can be reflected, the resource combination with high communication efficiency can be quickly screened, and the resource combinations can be screened from different resource pools to fully utilize resources in different states, improve the utilization rate of resources, and meet the performance demand of the computing task to speed up the response speed of the task.

[0063] During the computing task, the state of the computing resource changes, including but not limited to fluctuation of resource utilization, change of communication delay, dynamic coupling of the first resource combination and the second resource combination according to the real-time state change of the computing resource, which can adjust the allocation and cooperation mode of the computing resource according to the real-time state of the resource, and dynamically adjust the resource pool according to the change of resource utilization, including but not limited to expanding the resource pool, ensuring that the resource state in the resource pool can always reflect the corresponding performance, and improving the flexibility and effectiveness of computing resource management and scheduling. Dynamic coupling can dynamically adjust the resource combination according to the real-time state of the resource, optimize the resource cooperation mode, avoid resource waste and performance bottleneck, improve the processing efficiency of the computing task, and dynamically adjust the resource pool to ensure that the resource state in the resource pool is consistent with the real-time performance, provide accurate state basis for resource screening, and improve the effectiveness of resource management.

[0064] After the computing task is completed, the computing resources in the resource combination are decoupled, the association between the resources is removed, independent computing resources are obtained, the real-time state of the computing resources is re-evaluated, the computing resources are put into the corresponding resource pool according to the corresponding state, and the computing resources are managed in the pool. Decoupling and re-classification of resources after the task is completed can quickly restore the available state of the computing resources, improve the reuse rate of the resources, put the resources into the corresponding resource pool, facilitate resource screening and management for the next computing task, and realize sustainable pool management of heterogeneous computing resources.

[0065] The application maps the state vector of the heterogeneous resource to the resource double-chain structure for coding, puts the resources into different resource pools according to the resource coding, screens the corresponding resource combination from the resource pool according to the computing task in combination with the communication between the computing resources, dynamically couples the resources in the resource combination during the task computing, analyzes the resource utilization and resource coding difference to adjust the capacity of the resource pool, and decouples the computing resources after the computing task is completed and puts them into the corresponding resource pool; screening and dynamic coupling of the computing resources through resource coding in combination with the communication between the computing resources can quickly respond to the computing task of the user, make the heterogeneous computing resources be more fully utilized, reduce resource waste, and realize efficient pool management of the heterogeneous computing resources.

[0066] Further, the state of each computing resource in the resource pool is coded according to the real-time state data of the heterogeneous resource to obtain a resource code, including:

[0067] S201, state feature extraction is performed on the real-time state data of the heterogeneous resource to construct a state vector;

[0068] S202, the state vector is mapped to a pre-constructed resource double-chain structure for coding to obtain a resource code of each computing resource;

[0069] S203, put the computing resource with the resource code greater than the preset coding threshold into the first resource pool, and put the computing resource with the resource code less than or equal to the preset coding threshold into the second resource pool.

[0070] In the embodiment, according to the collected real-time state data of the heterogeneous computing resource, the state features related to the state are extracted, the state vector is constructed, the state vector is mapped to the pre-constructed resource double chain structure to code the resource, the resource double chain structure reflects the performance of the computing resource, according to the preset coding threshold, the computing resource with the resource code greater than the coding threshold is put into the first resource pool, and the computing resource with the resource code less than or equal to the coding threshold is put into the second resource pool, to realize the classified management of the resource; through the feature extraction and the resource double chain structure coding, the resource code accurately reflecting the state of the resource can be obtained, and the classified storage of the computing resource can select the corresponding resource combination according to different computing task requirements, to improve the resource utilization rate and the task response speed.

[0071] Specifically, the state features of the real-time state data of the heterogeneous resource are extracted, and the state vector is constructed; the real-time state data includes but is not limited to the usage rate, the cache hit rate, the total capacity of the memory, the used capacity, the state features are extracted from the real-time state data according to the demand of the resource performance evaluation, including the display memory occupancy rate, the computing power, the network receiving rate, etc., the extracted state features are standardized, and the standardized feature values are arranged in order to obtain the state vector of each computing resource; through the feature extraction, the data processing amount can be reduced, and the efficiency and accuracy of the resource coding can be improved.

[0072] Specifically, the state vector is mapped to the pre-constructed resource double chain structure for coding to obtain the resource code of each computing resource; the pre-constructed resource double chain structure is composed of a performance chain and a reliability chain, the performance chain reflects the performance indexes of the computing resource such as the computing power and the communication efficiency, and the reliability chain reflects the reliability indexes of the computing resource such as the stability and the failure probability, the state vector is mapped to the resource double chain structure, the state of the computing resource can be coded from the two dimensions of performance and reliability, the obtained resource code not only reflects the performance level of the resource, but also reflects the corresponding stability degree, to provide comprehensive data support for the selection and combination of the computing resource. The resource double chain structure is combined to code the computing resource from the two dimensions of performance and reliability, so that the state information in the resource code is more comprehensive, the limitation of single dimension evaluation is avoided, the state vector is mapped to the resource double chain structure to calculate the resource code, the corresponding resource combination can be selected according to the resource code, and the efficiency and accuracy of the resource combination and the pooling management are improved.

[0073] After the resource encoding is calculated, the computing resource with the resource encoding greater than the preset encoding threshold is put into the first resource pool, and the computing resource with the resource encoding less than or equal to the preset encoding threshold is put into the second resource pool; the preset encoding threshold is determined according to the resource demand of the historical computing task, the average performance level of the resource and the management experience, the resource encoding of each computing resource is compared with the preset encoding threshold, the computing resource with the resource encoding greater than the preset encoding threshold is put into the first resource pool, the resource in the first resource pool performs better in performance and reliability, and is suitable for the computing task with higher requirements on computing ability and stability; the computing resource with the resource encoding less than or equal to the preset encoding threshold is put into the second resource pool, and the resource in the second resource pool is suitable for the computing task with relatively lower performance requirement; the computing resource is divided according to the resource state and put into different resource pools for management, the resource can be quickly screened based on the computing task type, the efficiency of resource scheduling is improved, the resources in different resource pools can be differentially managed according to different task requirements, the value of various resources is fully utilized, resource mismatch and waste are avoided, and the resource utilization rate is improved.

[0074] Further, the state vector is mapped to a pre-constructed resource double chain structure for encoding to obtain the resource encoding of each computing resource, including:

[0075] S301, according to the dimension of the state vector, a feature space corresponding to the dimension is constructed, the state vector is mapped into the feature space to generate a feature sequence;

[0076] S302, combining the historical state data of the computing resource and the feature sequence, a resource double chain structure is constructed, the resource double chain structure includes a first resource chain constructed according to the feature sequence and a second resource chain constructed by complementary conversion of the feature sequence;

[0077] S303, based on the real-time resource load of the resource pool, the information entropy value of each computing resource at the corresponding position in the resource double chain structure is calculated to obtain the corresponding resource encoding.

[0078] In this embodiment, according to the dimension of the state vector, a feature space corresponding to the dimension is constructed, the state vector is mapped into the feature space to generate a feature sequence; each dimension of the state vector represents a state feature of the computing resource, a three-dimensional feature space is determined according to the dimension of the state vector, which provides a mapping space for the state vector, each dimension of the state vector is corresponded to the feature space, the coordinates of each state vector in the feature space are calculated to obtain the corresponding feature sequence; the state vector is mapped into the feature space, so that the state vector can be compared in the same feature space, the corresponding feature sequence is calculated, which is convenient for matching and integrating with the resource double chain structure, so as to calculate the resource encoding.

[0079] Specifically, the historical state data and the feature sequence of the computing resource are combined to construct a resource double-chain structure, the resource double-chain structure including a first resource chain constructed according to the feature sequence and a second resource chain constructed by complementary conversion of the feature sequence; the historical state data of the computing resource is standardized, and based on the feature sequence, similar feature sequences with a similarity greater than a preset similarity threshold are screened out according to the similarity between the historical feature sequence and the feature sequence, the similarity threshold being set according to the feature calculation accuracy requirement, the similar feature sequences are connected in time sequence to form the first resource chain, and the similar feature sequences are complementarily converted, each feature value of the feature sequence is normalized to 0 to 1, the absolute value of the difference between the feature value and 1 is calculated to obtain the corresponding complementary feature sequence, and the complementary feature sequence is connected in time sequence to form the second resource chain. The first resource chain reflects the continuity and similarity of the computing resource state according to the historical data, and can analyze the real-time change trend of the computing resource state; the second resource chain provides a supplementary perspective of the resource state through complementary conversion, and performs multi-dimensional analysis on the resource state, thereby improving the accuracy of the resource state coding result and providing accurate data support for the selection and dynamic coupling of the computing resource.

[0080] Specifically, the real-time resource load condition of the resource pool is analyzed, the resource load condition reflects the use pressure of the current resource pool resource, the information entropy value of each computing resource at the corresponding position on the resource double-chain structure is calculated to obtain the corresponding resource code; the resource pool load is calculated according to the real-time resource load data of the resource pool, the resource load data including but not limited to the average CPU utilization rate, the average memory occupancy rate and the task queue number of all resources in the resource pool, according to the position of the resource on the first resource chain and the second resource chain, each position is allocated a corresponding weight according to the proportion of the load of the computing resource in all resources, the higher the load, the greater the weight of the resource at the key position in the chain; according to the feature value distribution of each position of the computing resource on the resource double-chain structure and the corresponding weight, the information entropy value is calculated by using the information entropy calculation formula, and the information entropy calculation formula is as follows:

[0081] ;

[0082] In the formula, is the information entropy value of the computing resource at the position, is the probability of the feature value i in the feature sequence in the historical feature value distribution at the position, The number of feature values in the feature sequence is denoted as. After the information entropy value is calculated, the information entropy value is corrected according to the real-time load of the computing resource. The load of the computing resource is increased by 10%, and the information entropy value is multiplied by a correction coefficient of 1.1 to obtain the resource code of the computing resource. The information entropy value calculated based on the real-time load can reflect the comprehensive state of the computing resource under the current load environment, and the obtained resource code can reflect the real-time state of the computing resource. By quantifying the uncertainty of the resource state through the calculation of the information entropy value, different structures of the computing resource can be compared horizontally, so that the resource combination with high stability and load adaptation can be screened, and the effectiveness of the resource screening and scheduling process can be improved.

[0083] Further, in response to the computing task, a communication dependency graph is constructed by analyzing the communication between the computing resources through the preset resource combination model, a first resource combination is screened from the first resource pool, and a second resource combination is screened from the second resource pool, comprising:

[0084] S401, in response to the computing task, the communication dependency value between the computing resources is analyzed in combination with the resource code of the computing resource, a communication dependency graph and a task communication constraint are generated;

[0085] S402, through the preset resource combination model, resource nodes meeting the task communication constraint are screened from the communication dependency graph corresponding to the first resource pool to obtain a first resource combination, and resource nodes connected with the resource nodes in the first resource combination are screened from the communication dependency graph corresponding to the second resource pool to obtain a second resource combination;

[0086] S403, according to the first resource combination and the second resource combination screened, the communication dependency graph is updated in real time.

[0087] In this embodiment, when the computing task arrives, the communication dependency value between the computing resources is analyzed in combination with the resource code of the computing resource, a communication dependency graph and a task communication constraint are generated; the communication efficiency between different computing resources directly affects the execution speed of the computing task, the resource code reflects the performance, stability and real-time state of the computing resource, the communication dependency value is analyzed in combination with the resource code, the communication correlation degree between the computing resources is quantified, and the communication dependency graph is constructed to directly reflect the communication relationship between the resources. The corresponding task communication constraint is obtained by analyzing the computing task, and the minimum requirement of the computing task for resource communication is determined; the communication dependency value between the computing resources is analyzed in combination with the resource code, the comprehensive state of the resource is considered, the accuracy of the analysis result of the computing resource communication is improved, the communication dependency graph can reflect the communication relationship between the computing resources, the task communication constraint determines the screening condition of the computing task, provides accurate data support for the screening of the resource combination, and ensures that the screened resource combination can meet the communication demand of the task.

[0088] Specifically, by a preset resource combination model, resource nodes meeting the task communication constraints are filtered out from the communication dependency graph corresponding to the first resource pool, and the resource nodes meeting the core communication demand are preferentially filtered to obtain a first resource combination. Resources in the second resource pool are supplemented, and resource nodes connected to the resource nodes in the first resource combination are filtered out from the communication dependency graph corresponding to the second resource pool to obtain a second resource combination. The two resource combinations can cooperatively and efficiently complete the computing task. The resource combination model includes but is not limited to a particle swarm optimization model. The communication dependency graph and the task communication constraint are input into the particle swarm optimization model. The model preferentially selects a resource combination with a high communication dependency value, a low load, and meeting the constraint condition according to the task communication constraint to obtain the first resource combination. Based on the resource nodes in the first resource combination, resource nodes connected thereto are filtered out from the corresponding second resource pool resource nodes in the communication dependency graph to obtain the second resource combination. The first resource combination guarantees the high-performance communication demand of the core part of the task, and the second resource combination is closely related to the first resource combination in communication, thereby enhancing the cooperativity between the resource combinations and being conducive to improving the processing efficiency of the computing task.

[0089] After the resource combination is filtered out, the corresponding computing resources are allocated to execute the corresponding task, and the corresponding resource communication state will change accordingly, including but not limited to an increase in communication load and a fixed communication path. According to the communication state changes of the computing resources in the first resource combination and the second resource combination, the communication dependency values between the resource nodes are updated, and the communication dependency graph is updated in real time. Real-time updating of the communication dependency graph can reflect the actual communication state of the resource combination and improve the timeliness and accuracy of the resource pooling management.

[0090] Further, in response to the computing task, the communication dependency values between the computing resources are analyzed in combination with the resource encoding of the computing resources, the communication dependency graph and the task communication constraint are generated, including:

[0091] S501, according to the resource encoding of the computing resources and the topological structure between the resource nodes, a first communication value between the computing resources is calculated, and the first communication value is corrected by a stability coefficient in the resource encoding to obtain a communication dependency value;

[0092] S502, an associated edge is established between resource nodes with a communication dependency value greater than a preset dependency threshold, and a communication dependency graph is constructed;

[0093] S503, in response to the computing task, the dependency between the resources required by the task is analyzed, and a task communication constraint is generated in combination with the communication mode corresponding to the task type.

[0094] In this embodiment, the resource coding includes information such as the performance and load of the computing resources. The topology between resource nodes determines the basic conditions of the physical communication path, including distance and connection method. Both the resource coding and the topology affect the communication effect between resources. By combining the resource coding of the computing resources and the topology between resource nodes, the first communication value between resources is calculated, and the first communication value is corrected by the stability coefficient in the resource coding to obtain the communication dependency value.

[0095] Specifically, communication-related features, including network performance components and load state components, are extracted from the resource codes of computing resources. Topological information between resource nodes, including physical location, link type, and link length, is also obtained. The extracted topological information is assigned values ​​according to corresponding scoring rules, and each feature value is summed to obtain the first communication value. The resource codes include stability coefficients, reflecting hardware stability and software failure rate; higher stability coefficients indicate more stable resources. The first communication value is weighted using the average stability coefficients of two computing resources to calculate the communication dependency value between the two resources. Combining resource codes and topological structure in calculating the first communication value integrates the communication capabilities and physical connection conditions of the computing resources. Corrected by stability coefficients, the calculated communication dependency value better reflects the actual reliability of resource communication, improving the accuracy of the communication dependency value.

[0096] like Figure 2 As shown, an association edge is established between resource nodes whose communication dependency value is greater than a preset dependency threshold to construct a communication dependency graph. Based on historical communication data and task execution experience, the critical value for effective communication between resource nodes is determined, and a dependency threshold is set. All computing resource node pairs are traversed, and the communication dependency value of each pair of nodes is compared with the preset dependency threshold. Node pairs with communication dependency values ​​greater than the dependency threshold are selected. Using computing resources as nodes, an association edge is established between the selected effective node pairs, with the edge weight being the corresponding communication dependency value, thus obtaining the communication dependency graph. By filtering effective node pairs through the preset dependency threshold, it is ensured that the association edges in the communication dependency graph reflect high-quality communication relationships. The constructed communication dependency graph can reflect the communication capabilities between computing resources, thereby quickly filtering out the corresponding resource combinations.

[0097] The subtask division and dependency relationship of different computing tasks are different, and the demand for resource communication is also different. In response to the computing task, the dependency between the required resources of the analysis task is analyzed, the data interaction demand between the subtasks is determined, the corresponding task communication constraint is obtained by combining the communication mode corresponding to the task type, including but not limited to point-to-point communication and broadcast communication. The structure of the computing task is analyzed, the number of subtasks, the dependency relationship between the subtasks and the data interaction amount are determined, and the task type and the corresponding communication mode are determined according to the application scene and the processing mode of the task. The task type includes but is not limited to distributed computing and real-time stream processing; according to the determined communication mode, the key feature parameters are extracted, including the maximum delay and bandwidth requirement of point-to-point communication, the synchronization accuracy and data consistency requirement of broadcast communication, etc.; the specific task communication constraint is generated by combining the task dependency and the communication mode feature parameters; by analyzing the task dependency and the communication mode, the generation of the task communication constraint is more in line with the actual demand of the computing task, the constraint condition provides a judgment standard for the resource combination screening, and it is ensured that the screened resource combination can meet the communication characteristics of the task, and the reliability and efficiency of the task execution are improved.

[0098] Further, in the process of the computing task, the first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resource, and the utilization rate of the computing resource is analyzed to dynamically adjust the first resource pool and the second resource pool, including:

[0099] S601, analyzing the real-time performance index of the task in the process of the computing task, and constructing a task performance vector;

[0100] S602, dynamically coupling the computing resource by combining the task performance vector and the communication dependency value of the computing resource in the first resource combination and the second resource combination, and obtaining a coupled resource;

[0101] S603, dynamically adjusting the first resource pool and the second resource pool by analyzing the utilization rate of the coupled resource and the resource code difference value of the computing resource.

[0102] The embodiment collects real-time performance indicators during the execution of a computing task, constructs a task performance vector after standardization, reflects the current execution state of the computing task, performs correlation analysis on the task performance vector and the communication dependency value of the resources in the resource combination, identifies performance bottlenecks and formulates adjustment strategies, dynamically couples the computing resources in the first resource combination and the second resource combination, optimizes the resource collaboration mode, calculates the comprehensive utilization rate and resource coding difference value of the coupled resources, and dynamically adjusts the first resource pool and the second resource pool to ensure that the state of the resource pool accurately reflects the actual performance of the computing resources. By constructing the task performance vector in real time and dynamically coupling the resources, the collaborative communication capability between the resources is enhanced, the resource combination can quickly respond to changes in task performance, timely solve performance bottlenecks, and improve the anti-interference capability and stability of the task in a complex environment; based on the utilization rate and the resource coding difference value, the resource pool is adjusted, so that the resource pool can reflect the performance change of the resources in real time, avoid resource mismatch, and provide accurate resource classification basis for resource selection of the computing task.

[0103] In the embodiment, the performance of the task fluctuates with the changes of the resource state and the load during the execution of the computing task. Analyzing the real-time performance indicators of the task can reflect the current execution efficiency and resource demand of the task, and a task performance vector is constructed. The real-time performance indicators include but are not limited to task progress completion rate, subtask response time, data transmission rate, calculation error rate, and resource utilization rate matching degree. The collected performance indicator data is standardized, and each standardized performance indicator is arranged in order to obtain the task performance vector. The collection and vector construction of the real-time performance indicators can timely capture the performance changes in the task execution process, provide accurate data basis for subsequent dynamic coupling of computing resources, and ensure that the resource adjustment can accurately match the real-time demand of the task.

[0104] Specifically, the computing resources are dynamically coupled to obtain coupled resources by combining the task performance vector and the communication dependency value of the computing resources in the first resource combination and the second resource combination. The task performance vector reflects the current execution state and performance bottleneck of the task, and the communication dependency value of the computing resources in the resource combination reflects the communication collaboration capability between the resources. Dynamic coupling of the computing resources based on the task performance vector and the communication dependency value can adjust the collaboration mode of the resources according to the task performance changes, optimize the communication path and task allocation between the resources, make the performance of the resource combination dynamically match the task demand, and improve the task execution efficiency.

[0105] Meanwhile, the utilization of the coupled resources and the resource encoding difference of the computing resources are analyzed, and the first resource pool and the second resource pool are dynamically adjusted. The utilization of the coupled resources reflects the actual load condition of the computing resources. The utilization that is too high or too low indicates that the resource configuration is unreasonable. The resource encoding difference of the computing resources reflects the difference degree of the resource performance. The resource pool is dynamically adjusted in combination with the utilization and the resource encoding difference, so that the resources with performance changes can be classified into appropriate resource pools, the resource state in the resource pool is ensured to be consistent with the actual performance, and the accuracy of the resource pool and the resource scheduling efficiency are improved.

[0106] Further, the computing resources are dynamically coupled in combination with the task performance vector and the communication dependency values of the computing resources in the first resource combination and the second resource combination, to obtain coupled resources, including:

[0107] S701, in combination with the task performance vector, the real-time state of the computing resources, and the communication dependency values of the computing resources in the first resource combination and the second resource combination, the coupling weights between the corresponding computing resources are calculated.

[0108] S702, the computing resources with coupling weights greater than a preset coupling threshold are screened out in the first resource combination, to obtain first coupled resources.

[0109] S703, the computing resources with coupling weights greater than the preset coupling threshold between the first coupled resources are screened out in the second resource combination, to obtain second coupled resources.

[0110] S704, the first coupled resources are coupled with the second coupled resources, to obtain coupled resources.

[0111] In the embodiment, the task performance vector reflects the current execution state and demand priority of the computing task, the real-time state of the computing resources reflects the current load and availability of the computing resources, and the communication dependency value reflects the communication and cooperation capability between the computing resources. In combination with the task performance vector, the real-time state of the computing resources, and the communication dependency values of the computing resources in the first resource combination and the second resource combination, the adaptation degree between the computing resources is quantified by comprehensively considering the computing task demand, the resource state and the communication quality, the coupling weights between the corresponding computing resources are calculated, and data support is provided for resource screening and coupling.

[0112] Specifically, according to the task type and the resource management target, corresponding total weights are assigned to the task performance vector, the real-time state of the computing resource and the communication dependency value, the parameters related to resource coupling are extracted from the task performance vector, including the subtask response time standardized value, the data transmission rate standardized value and the like, and corresponding sub-weights are assigned to the parameters according to the influence degree of the parameters on the coupling; the real-time state of the computing resource is standardized, and the real-time state includes but is not limited to the CPU utilization rate, the memory utilization rate and the load rate, and corresponding sub-weights are assigned to the state parameters; the task performance vector parameters, the resource real-time state quantized values and the communication dependency values are weighted and summed with the corresponding sub-weights, and the calculated values are weighted and summed with the corresponding total weights respectively to obtain the coupling weight. The coupling weight is calculated by comprehensively considering the task performance, the resource state and the communication dependency value, and the calculated coupling weight result can accurately reflect the adaptation degree of the resource and the task, avoid the unreasonable resource coupling caused by the single factor consideration, and provide accurate data support for resource screening.

[0113] As shown in Figure 3 , the resource performance in the first resource combination is better, the computing resources with the coupling weight greater than the preset coupling threshold value are screened out in the first resource combination first to obtain the first coupling resource, so as to ensure the adaptability and the collaborative ability between the first coupling resources and provide guarantee for the efficient execution of the task; the coupling threshold value is set according to the demand of the task for the computing resource, all the computing resources in the first resource combination are traversed, the coupling weight of each resource and other resources in the combination is obtained, the resource pairs with the coupling weight greater than the preset coupling threshold value are screened out to obtain the first coupling resource; the first coupling resource is screened through the coupling threshold value, so that the high adaptability and the collaborative ability between the core resources can be ensured, the execution efficiency of the computing task is improved, and the performance bottleneck caused by the poor cooperation of the core resources is reduced.

[0114] Correspondingly, the resources in the second resource combination serve as the supplementary resources, need to maintain good collaborative communication with the first coupling resources, the computing resources with the coupling weight greater than the preset coupling threshold value between the first coupling resources are screened out in the second resource combination to obtain the second coupling resource; the smooth communication and the efficient cooperation between the supplementary resources and the core resources can be ensured, and the overall optimization of the resource combination is realized. The coupling weight of each resource in the second resource combination and each resource in the first coupling resource is calculated, the average value of the coupling weight of each resource in the first coupling resource is taken as the overall coupling weight of the resource and the first coupling resource; the second resource combination resources with the overall coupling weight greater than the preset coupling threshold value are screened out as the second coupling resource. The second coupling resource with high coupling degree with the first coupling resource is screened out, so that the supplementary resources and the core resources can efficiently cooperate, the overall performance decline caused by the poor adaptability between the supplementary resources and the core resources is avoided, and the collaborative efficiency of the resource combination is improved.

[0115] The first coupling resource is coupled with the second coupling resource, the advantages of the core resource and the supplementary resource are integrated, and the coupling resource is obtained. A corresponding communication connection is established between the first coupling resource and the second coupling resource, a communication protocol and a data transmission rule are configured according to a resource type and a state, data between resources can be smoothly transmitted, a core submodule of a computing task is allocated to the first coupling resource according to a resource performance and a coupling weight, and a supplementary submodule is allocated to the second coupling resource; and a task interaction mode, a data transmission frequency and a synchronization mechanism between the first coupling resource and the second coupling resource are determined, the first coupling resource, the second coupling resource and the communication connection and the cooperation rule thereof are integrated, and the complete coupling resource is obtained. The first coupling resource and the second coupling resource are coupled, the advantages of the core resource and the supplementary resource are complementary, the complete and efficient resource cooperation system is obtained, the communication connection and the cooperation rule are determined, the smooth cooperation between resources can be ensured, and the overall execution efficiency of the computing task is improved.

[0116] Further, the utilization rate of the coupling resource and the resource code difference value of the computing resource are analyzed, and the first resource pool and the second resource pool are dynamically adjusted, including:

[0117] S801, the utilization rate of the coupling resource is analyzed, if the utilization rate of the first coupling resource in the coupling resource is greater than a preset utilization rate threshold, an expansion resource with the same structure as the computing resource in the first coupling resource and the highest resource code is selected from the second resource pool;

[0118] S802, the expansion resource is put into the first resource pool, the first resource pool is expanded, and the second resource pool is updated;

[0119] S803, if the absolute value of the continuous resource code difference value in the second resource pool is greater than a preset code difference threshold, the computing resource with the low resource code in the continuous computing resource is put into the cooling resource area in the second resource pool.

[0120] In the embodiment, the utilization rate of the coupling resource is analyzed, the first coupling resource in the coupling resource is used as the core resource, the utilization rate of the core resource is too high, the load pressure of the core resource is large, and the task execution efficiency is affected, if the utilization rate of the first coupling resource in the coupling resource is greater than a preset utilization rate threshold, an expansion resource with the same structure as the computing resource in the first coupling resource and the highest resource code is selected from the second resource pool, and the resource pool is adjusted.

[0121] According to the performance characteristics of the first coupling resource and the task type, a utilization threshold is set, the utilization data of the first coupling resource in the coupling resource is continuously collected, including CPU utilization, memory utilization, etc., the average value of the utilization is taken to calculate the comprehensive utilization; the comprehensive utilization of the first coupling resource is compared with the utilization threshold, if it is greater than the utilization threshold, the same type of homogeneous resource in the second resource pool is screened, in the homogeneous resource, the resource with the highest resource code is selected as the expansion resource. When the core resource is overloaded, the homogeneous high code resource is supplemented in time, which can quickly relieve the pressure of the core resource, avoid the performance degradation caused by high load, the homogeneous resource reduces the complexity of resource adaptation, the high code resource guarantees the performance of the supplemented resource, and ensures the continuous and efficient execution of the task.

[0122] As shown in Figure 4 The expansion resource is put into the first resource pool, the first resource pool is expanded, the processing capacity of the core resource is enhanced, the second resource pool is updated, the expanded expansion resource is removed, the accuracy of the state of the second resource pool is ensured; the screened expansion resource is migrated from the second resource pool to the first resource pool, the information of the expansion resource is added in the resource list of the first resource pool, including resource code, type, real-time state, etc., the expansion is completed, the migrated expansion resource information is deleted from the resource list of the second resource pool, and the accuracy of the second resource pool data is ensured. The expansion of the first resource pool enhances the processing capacity of the core resource, which can better cope with the high load task demand, the second resource pool is updated in time to ensure the accuracy of the resource information, which provides a reliable basis for subsequent resource screening and management, and improves the dynamic adaptability of the resource pooling management.

[0123] Meanwhile, the resource code difference in the second resource pool reflects the difference degree of resource performance, and the continuous large resource code difference indicates that the resource performance is uneven. If the absolute value of the continuous resource code difference in the second resource pool is greater than the preset code difference threshold, the computing resource with low resource code in the continuous computing resource is put into the cooling resource area in the second resource pool, the poor performance resource is isolated, and the influence on the resource screening efficiency is avoided. According to the resource performance distribution of the second resource pool, the code difference threshold is set, the resources in the second resource pool are sorted from high to low according to the resource code, the code difference between adjacent resources is calculated, when the absolute value of the continuous resource code difference is greater than the preset code difference threshold, the computing resource with low resource code is determined as the cooling resource, the cooling resource area is divided in the second resource pool, the screened cooling resource is migrated to the area, the cooling resource is suspended to participate in the resource screening, the state of the computing resource in the cooling resource area is monitored and maintained, and after the state is improved, the resource code can be re-evaluated and participate in the resource screening. The low code resource with large performance difference in the continuous computing resource is put into the cooling resource area, which can reduce the influence of the low performance computing resource in the second resource pool on the screening efficiency, improve the precision of the computing resource screening, provide space for resource maintenance for the cooling resource area, help the recovery and reuse of the resource state, and optimize the resource quality of the second resource pool.

[0124] Further, after the computing task is completed, the first resource combination and the second resource combination are decoupled, and according to the state of the decoupled computing resource, the corresponding resource pool is put into, so as to pool the heterogeneous computing resources, including:

[0125] S901, after the computing task is completed, the state of the computing resource in the first resource combination and the second resource combination is analyzed, and the corresponding decoupling degree is calculated;

[0126] S902, according to the decoupling degree from high to low, the computing resource is decoupled to obtain the decoupled resource;

[0127] S903, the state of the decoupled resource is analyzed, the corresponding resource code is calculated, and the decoupled resource is put into the corresponding resource pool according to the resource code.

[0128] In this embodiment, after the completion of the computing task, the computing resources in the first resource combination and the second resource combination still have associated relationships such as communication connection, task dependency, and the like. The state of the computing resources in the first resource combination and the second resource combination is analyzed, and the corresponding decoupling degree is calculated. After the completion of the computing task, the state data of the computing resources in the first resource combination and the second resource combination is collected, including the current communication connection list, the amount of incomplete data transmission, the task residual dependency mark, and the like. The collected state data is standardized to the 0-1 interval, wherein the indicators such as the number of communication connections and the amount of data sharing that are negatively correlated with the decoupling degree are inversely standardized. The sum of the state data values is calculated, and the corresponding decoupling degree is calculated. The decoupling degree quantifies the degree of association between resources. The higher the decoupling degree, the weaker the dependency between resources, and the easier the separation. The lower the decoupling degree, the closer the association. By calculating the decoupling degree, the degree of association between resources can be quantified, and the blindness of the decoupling operation can be avoided. The calculated decoupling degree provides a sequence standard for decoupling by priority, ensures that resources with weak association are decoupled first, and reduces the impact of the decoupling process on the resource state.

[0129] According to the decoupling degree from high to low, the computing resources are decoupled, and resources with weak association are preferentially separated to reduce conflicts and risks in the decoupling process. Resources with weak association have a small decoupling difficulty and are decoupled first to quickly release part of the resources. Resources with strong association are decoupled later, which has sufficient time to handle complex dependency relationships, ensures complete decoupling, and does not damage the resource state, thereby obtaining decoupled resources. The decoupling degrees of all resource pairs in the first resource combination and the second resource combination are sorted from high to low. According to the sorting result, the order of resource decoupling is determined, the resource pair with the highest decoupling degree is preferentially decoupled, and the resources are sequentially decoupled according to the decoupling order. The decoupling operation includes closing communication connections, clearing data sharing links, removing task dependency relationships, releasing cooperative locks, and the like. After each group of resources is decoupled, it is checked whether the resources have been completely separated and whether there are residual dependencies or connections, thereby obtaining decoupled resources. Decoupling according to the decoupling degree sorting reduces the error probability of decoupling complex associated resources, ensures a smooth and efficient decoupling process, preferentially releases resources with weak association, quickly increases the number of available resources in the resource pool, and improves resource reuse efficiency. Decoupling verification ensures complete resource separation, avoids residual dependencies affecting subsequent resource use, and improves the accuracy and effectiveness of computing resource pooling management.

[0130] A heterogeneous computing resource pooling management system for implementing a heterogeneous computing resource pooling management method, comprising:

[0131] A resource encoding module encodes the state of each computing resource in the resource pool according to real-time state data of the heterogeneous resources to obtain resource encoding.

[0132] The resource screening module screens out a first resource combination from the first resource pool and a second resource combination from the second resource pool by analyzing communication conditions between the computing resources through a preset resource combination model in response to the computing task.

[0133] The resource coupling module dynamically couples the first resource combination and the second resource combination according to real-time states of the computing resources during execution of the computing task, and dynamically adjusts the first resource pool and the second resource pool according to utilization rates of the computing resources.

[0134] The resource decoupling module decouples the first resource combination and the second resource combination after the computing task ends, and puts the decoupled computing resources into the corresponding resource pool according to states of the decoupled computing resources, so as to pool manage the heterogeneous computing resources.

[0135] In the embodiment, the resource coding module codes each resource by collecting real-time state data of the heterogeneous computing resources, including but not limited to CPU utilization, memory occupancy, and network bandwidth, performing feature extraction, standardization processing, and coding calculation on the state of each resource, obtaining resource codes reflecting performance and stability of the resources, and dividing the resources into the first resource pool and the second resource pool based on a preset coding threshold. The resource state is analyzed and the resources are classified and stored, which provides data support for resource screening and pool management, ensures the accuracy of resource state evaluation, and improves the efficiency of resource pool management.

[0136] The resource screening module analyzes communication conditions between the resources through a preset resource combination model after receiving the computing task, constructs a communication dependency graph, generates constraint conditions in combination with communication requirements of the task, screens out a first resource combination meeting core performance and communication constraints from the first resource pool, and screens out a second resource combination having high communication synergy with the first resource combination from the second resource pool. The corresponding resource combination can be matched through communication dependency analysis, the advantages of different resource pools are fully brought into play, the selected resource combination can efficiently respond to task requirements, resource mismatch risks are reduced, and task startup speed is accelerated.

[0137] The resource coupling module analyzes resource states and task performance indicators of the first resource combination and the second resource combination in real time during execution of the computing task, dynamically adjusts resource cooperation modes through coupling weights between the computing resources, optimizes communication paths and task allocation between the resources, dynamically couples the computing resources, dynamically adjusts the resource pools according to changes in resource utilization and resource codes, migrates resources with changed performance between the two resource pools, dynamically optimizes the resource combination, timely responds to performance fluctuations and resource load changes during task execution, avoids resource bottlenecks, improves resource utilization and task processing efficiency, and enhances flexibility and adaptability of resource management.

[0138] The resource decoupling module calculates the decoupling degree by analyzing the communication connection, data sharing and other association relationships between resources after the end of the computing task, sequentially releases the dependency relationship between the resource combinations according to the decoupling degree from high to low, obtains independent decoupled resources, re-evaluates the state of the decoupled resources and calculates the resource code, puts the decoupled resources back into the corresponding resource pool, can ensure that the resources are completely released and the state is complete, realizes the recycling of the computing resources through the re-encoding, ensures that the resource pool state is synchronized with the actual performance of the resources, improves the resource reuse rate and the response speed of the computing task.

[0139] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the idea of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary technical personnel in the technical field, some improvements and refinements without departing from the principles of the present application shall also be considered as the protection scope of the present application.

Claims

1. A heterogeneous computing resource pooling management method, characterized in that, The method comprises the following steps: According to the real-time state data of the heterogeneous resources, the state of each computing resource in the resource pool is encoded to obtain a resource code, wherein the resource pool comprises a first resource pool storing resource codes greater than a preset encoding threshold and a second resource pool storing resource codes less than or equal to the preset encoding threshold; In response to a computing task, the communication dependency values between the computing resources are analyzed in combination with the resource codes of the computing resources to generate a communication dependency graph and a task communication constraint; Through a preset resource combination model, resource nodes satisfying the task communication constraint are filtered out from the communication dependency graph corresponding to the first resource pool to obtain a first resource combination, and resource nodes connected to the resource nodes in the first resource combination are filtered out from the communication dependency graph corresponding to the second resource pool to obtain a second resource combination; The communication dependency graph is updated in real time according to the filtered first and second resource combinations; During the process of the computing task, the first and second resource combinations are dynamically coupled according to the real-time state of the computing resources, and the utilization rate of the computing resources is analyzed to dynamically adjust the first and second resource pools; After the computing task is completed, the first and second resource combinations are decoupled, and the computing resources after decoupling are put into the corresponding resource pool for pool management of the heterogeneous computing resources; The method comprises the following steps: State feature extraction is performed on the real-time state data of the heterogeneous resources to construct a state vector; According to the dimension of the state vector, a feature space corresponding to the dimension is constructed, the state vector is mapped into the feature space, and a feature sequence is generated; A resource double-chain structure is constructed in combination with the historical state data of the computing resources and the feature sequence, wherein the resource double-chain structure comprises a first resource chain constructed according to the feature sequence and a second resource chain constructed by complementary conversion of the feature sequence; Based on the real-time resource load of the resource pool, the information entropy value of each computing resource at the corresponding position on the resource double-chain structure is calculated to obtain a corresponding resource code; Computing resources with a resource code greater than a preset encoding threshold are put into the first resource pool, and computing resources with a resource code less than or equal to the preset encoding threshold are put into the second resource pool.

2. The heterogeneous computing resource pooling management method of claim 1, wherein, The method comprises the following steps: According to the resource codes of the computing resources and the topological structure between the resource nodes, a first communication value between the computing resources is calculated, the first communication value is modified by a stability coefficient in the resource code to obtain a communication dependency value; An association edge is established between resource nodes with a communication dependency value greater than a preset dependency threshold to construct a communication dependency graph; In response to a computing task, the dependency between the resources corresponding to the task is analyzed, and a task communication constraint is generated in combination with the communication mode corresponding to the task type.

3. The heterogeneous computing resource pooling management method of claim 1, wherein, The first resource combination and the second resource combination are dynamically coupled according to the real-time state of the computing resources during the computing task, and the utilization of the computing resources is analyzed to dynamically adjust the first resource pool and the second resource pool, comprising: During the computing task, the real-time performance indicators of the task are analyzed to construct a task performance vector; The computing resources are dynamically coupled in combination with the task performance vector and the communication dependency values of the computing resources in the first resource combination and the second resource combination, to obtain coupled resources; The utilization of the coupled resources and the resource coding difference values of the computing resources are analyzed to dynamically adjust the first resource pool and the second resource pool.

4. The heterogeneous computing resource pooling management method of claim 3, wherein, The computing resources are dynamically coupled in combination with the task performance vector and the communication dependency values of the computing resources in the first resource combination and the second resource combination, to obtain coupled resources, comprising: The coupling weights between the corresponding computing resources are calculated in combination with the task performance vector, the real-time state of the computing resources, and the communication dependency values of the computing resources in the first resource combination and the second resource combination; The computing resources with coupling weights greater than a preset coupling threshold in the first resource combination are screened to obtain first coupled resources; The computing resources with coupling weights greater than the preset coupling threshold between the first coupled resources and the second resource combination are screened to obtain second coupled resources; The first coupled resources and the second coupled resources are coupled to obtain coupled resources.

5. The heterogeneous computing resource pooling management method of claim 3, wherein, The utilization of the coupled resources and the resource coding difference values of the computing resources are analyzed to dynamically adjust the first resource pool and the second resource pool, comprising: The utilization of the coupled resources is analyzed, and if the utilization of the first coupled resources in the coupled resources is greater than a preset utilization threshold, an extended resource with the same structure as the computing resources in the first coupled resources and the highest resource coding is selected from the second resource pool; The extended resource is placed in the first resource pool to expand the first resource pool and update the second resource pool; If the absolute value of the continuous resource coding difference values in the second resource pool is greater than a preset coding difference threshold, the computing resource with the lower resource coding in the continuous computing resources is placed in the cooling resource area in the second resource pool.

6. The heterogeneous computing resource pooling management method of claim 1, wherein, After the computing task is completed, the first resource combination and the second resource combination are decoupled, and the decoupled computing resources are placed in the corresponding resource pools according to the state of the decoupled computing resources, to pool the heterogeneous computing resources, comprising: After the computing task is completed, the state of the computing resources in the first resource combination and the second resource combination is analyzed to calculate the corresponding decoupling degree; The computing resources are decoupled from high to low according to the decoupling degree to obtain decoupled resources; The state of the decoupled resources is analyzed to calculate the corresponding resource coding, and the decoupled resources are placed in the corresponding resource pools according to the resource coding.

7. A heterogeneous computing resource pooling management system, characterized by, A heterogeneous computing resource pooling management method for realizing any one of claims 1-6, comprising: A resource coding module encodes the state of each computing resource in the resource pool according to the real-time state data of the heterogeneous resources to obtain resource coding. The resource screening module, in response to the computing task, analyzes communication conditions between computing resources through a preset resource combination model to construct a communication dependency graph, screens a first resource combination from a first resource pool, and screens a second resource combination from a second resource pool; The resource coupling module, in the process of the computing task, dynamically couples the first resource combination and the second resource combination according to real-time states of the computing resources, analyzes utilization rates of the computing resources, and dynamically adjusts the first resource pool and the second resource pool; The resource decoupling module, after the computing task ends, decouples the first resource combination and the second resource combination, and puts the decoupled computing resources into corresponding resource pools according to states of the decoupled computing resources, to perform pool management on the heterogeneous computing resources.

Images