Storage and calculation training AI server energy efficiency optimization method based on space-time deep cooperative game

Abstract

The invention discloses a storage and calculation training AI server energy efficiency optimization method based on a space-time deep cooperative game. The method comprises the following steps: S1, obtaining a multi-task load information set of an AI server; s2, initializing a space-time deep collaborative game model based on the multi-task load information set; s3, arranging an intelligent cache unit on a path for transmitting data from the storage unit to the calculation unit and returning the data from the calculation unit to the storage unit; s4, generating a dynamic caching strategy; s5, optimizing a two-way transmission path of the data; s6, the cache position and path of the high-frequency access data are redistributed, so that the cache strategy adapts to the optimal data flow direction under the current load condition; s7, optimizing the energy efficiency performance of the data from the storage unit to the calculation unit and a return path; and S8, further analyzing a resource competition state between different task loads in the AI server under the action of a feedback adjustment mechanism of the space-time deep cooperative game model. According to the method, the professional and technical thresholds required by traditional modeling are reduced, and the universality and convenience of three-dimensional modeling are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AI servers, and in particular to a method for optimizing energy efficiency of storage, computing and training AI servers based on deep spatiotemporal collaborative game. Background Art

[0002] With the development of artificial intelligence technology and high-performance computing, AI servers have become the core infrastructure for large-scale data processing and complex task computing. AI servers are widely used in deep learning, natural language processing, and computer vision. The performance of AI servers directly affects the efficiency of model training and reasoning. However, with the increase in task scale and complexity, the energy efficiency management and resource optimization issues of AI servers have become increasingly prominent.

[0003] At present, traditional AI server resource allocation and scheduling methods mainly rely on static configuration or scheduling strategies based on simple rules. Traditional methods are usually designed for fixed task loads and resource requirements, and are difficult to adapt to multi-task dynamic load environments. The allocation of computing resources and storage resources in traditional systems is often carried out independently, lacking consideration of the synergistic relationship between the two, resulting in inefficient resource allocation. In addition, due to the frequent transmission of data between storage units and computing units, traditional methods lack optimization strategies for data transmission paths, which further increases the system's resource consumption and energy overhead.

[0004] In the existing technology, some studies have attempted to optimize server resources by introducing load balancing and resource monitoring mechanisms. However, these methods usually only focus on the optimization of a certain type of resources, such as improving the utilization of computing units or optimizing the layout of data storage, but fail to comprehensively consider the synergistic relationship between storage, computing, and data transmission from a global perspective. In addition, the existing technology lacks the ability to cope with resource competition and dynamic adjustment in a multi-task load environment, and it is difficult to handle complex resource allocation conflicts between multiple tasks.

[0005] More importantly, traditional resource scheduling methods usually ignore the dynamic changes of tasks in the time and space dimensions. The data access patterns and execution durations of different tasks have significant time and space characteristics, and these characteristics have not been effectively utilized to optimize data transmission paths and cache strategies, resulting in high-frequency access data occupying too many resources, while low-frequency access data performs worse in terms of resource utilization and energy efficiency.

[0006] In summary, the existing technologies have significant defects in resource allocation efficiency, spatiotemporal data transmission optimization, and global resource competition processing under multi-task loads, and are difficult to meet the high requirements of modern AI servers for energy efficiency optimization. To address these problems, a new method is urgently needed to improve the resource utilization efficiency and energy efficiency performance of AI servers in dynamic load environments. Summary of the invention

[0007] One purpose of the present invention is to propose a method for optimizing the energy efficiency of storage, computing and training AI servers based on deep spatiotemporal collaborative game. The present invention reduces the professionalism and technical threshold required for traditional modeling and improves the universality and convenience of three-dimensional modeling.

[0008] According to an embodiment of the present invention, a method for optimizing energy efficiency of a storage, computing and training AI server based on time-space deep collaborative game includes the following:

[0009] S1. Obtain a multi-task load information set of the AI ​​server, including resource requirements of parallel tasks, data transmission requirements, and computing resource usage;

[0010] S2. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set. The spatiotemporal deep collaborative game model includes multiple agents for representing the data flow between the storage unit and the computing unit in the AI ​​server. Each agent corresponds to a different storage unit or computing unit to represent its own data storage and transmission requirements.

[0011] S3. An intelligent cache unit is set on the path where data is transmitted from the storage unit to the computing unit and returned from the computing unit to the storage unit. The intelligent cache unit analyzes the data access frequency between the storage unit and the computing unit based on the spatiotemporal depth collaborative game model, dynamically adjusts the storage location and cache strategy of the cached data, and realizes intelligent data caching on the bidirectional transmission path;

[0012] S4. Analyze the access frequency and transmission requirements of different tasks on the data transmission path based on the spatiotemporal deep collaborative game model, and generate a dynamic cache strategy, which includes the intelligent cache unit adjusting the cache position of high-frequency access data and low-frequency access data on different transmission paths;

[0013] S5. According to the dynamic cache strategy, high-frequency access data is preferentially cached in the cache unit close to the computing unit, and low-frequency access data is pre-stored in the cache unit close to the storage unit to optimize the bidirectional data transmission path;

[0014] S6. When the task load changes, the new task load information and its resource demand changes for storage units and computing units are analyzed in real time through the spatiotemporal deep collaborative game model, and the cache strategy of the intelligent cache unit is updated based on the analysis results, and the cache location and path of high-frequency access data are reallocated to adapt the cache strategy to the optimal data flow under the current load conditions;

[0015] S7. Automatically release redundant cache data in the intelligent cache unit after executing the cache strategy update, and monitor the load of the data transmission path in real time through the feedback adjustment mechanism of the spatiotemporal deep collaborative game model to optimize the energy efficiency performance of the data transmission path from the storage unit to the computing unit and the return path;

[0016] S8. Under the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, the resource competition status between different task loads in the AI ​​server is further analyzed, and the global optimal resource allocation strategy is generated in terms of task priority, cache unit location and path selection to improve the overall energy efficiency of the AI ​​server to the optimal state.

[0017] Optionally, the S1 includes:

[0018] S11, obtaining resource requirement information of each parallel task in the AI ​​server, wherein the resource requirement information includes the computing resource requirement C of each task i , Storage resource requirements M i And data transmission requirements D i , where i is the task identifier, i=1, 2, ..., N, and N is the total number of current parallel tasks;

[0019] S12. Constructing multi-task load information set:

[0020] L={(C i , M i , D i )|i=1,2,...,N}.

[0021] Optionally, S2 includes:

[0022] S21. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set L and construct a multi-agent set A = {a j |j=1,2,...,K}, where K is the total number of agents, and each agent a j Corresponds to a storage unit or computing unit in the AI ​​server, used to represent its data storage and transmission requirements at time t and spatial location s;

[0023] S22. For each agent a j Define the strategy space S j , its strategies j (t,s)∈S j Represents agent a j Resource allocation scheme at time t and spatial location s:

[0024] Computational resource allocationc j (t, s) represents agent a j The amount of computing resources allocated at time t and location s;

[0025] Storage resource allocation j (t, s) represents agent a j The amount of storage resources allocated at time t and location s;

[0026] Data transmission bandwidth allocationb j (t, s) represents agent a j The data transmission bandwidth allocated at time t and location s;

[0027] S23. Define the resource requirement function for each task i:

[0028] D i (t, s) = (C i (t,s),M i (t,s),B i (t, s));

[0029] Among them, C i (t, s) represents the computing resource requirements of task i at time t and location s, M i (t, s) represents the storage resource requirement of task i at time t and location s, B i (t, s) represents the data transmission bandwidth requirement of task i at time t and location s;

[0030] S24, establish a resource matching matrix R between tasks and agents to represent the resource allocation relationship between tasks and agents in the spatiotemporal dimension ij (t, s):

[0031]

[0032] S25. Define agent a j The utility function U j (t, s), used to measure its resource utilization efficiency:

[0033]

[0034] Among them, N is the total number of tasks, α1, β1, γ1 are weight coefficients, indicating the importance of computing resources, storage resources and transmission bandwidth, δ1 is the energy penalty coefficient, which is used to balance the relationship between resource utilization efficiency and energy consumption, E j (t, s) is the agent a j Energy consumption at time t and location s:

[0035] E j (t, s) = η c ·c j (t, s)+η m ·m j(t, s)+η b b j (t, s);

[0036] Among them, η c , η m , η b are the energy consumption coefficients of computing resources, storage resources, and transmission bandwidth respectively;

[0037] S26. Determine the optimal strategy for each agent by solving the optimization problem To maximize the total system utility:

[0038]

[0039] Among them, the constraints are task resource requirements constraints:

[0040]

[0041]

[0042]

[0043] Agent resource capacity constraints:

[0044]

[0045]

[0046]

[0047]

[0048] in, They are respectively j The maximum computing resources, storage resources and data transmission bandwidth capacity, s j (t, s) is the agent a j The strategy includes the resource allocation scheme at time t and location s.

[0049] Optionally, the S3 includes:

[0050] S31, set a smart cache unit set C = {c k |k=1,2,...,K}, where K is the total number of cache units, and the smart cache unit c k An intermediate node located in the transmission path, used to temporarily store data from the storage unit to the computing unit and from the computing unit back to the storage unit;

[0051] S32. Define each cache unit c kThe initial state variables are:

[0052]

[0053] in, Represents cache unit c k The currently available free storage capacity, A k Represents cache unit c k The data access frequency matrix of the internal cache, the element A of the matrix ij,k Indicates that the data of task i is in node c k The frequency of being visited by task j;

[0054] S33, based on the spatiotemporal depth collaborative game model, calculate the access frequency matrix F between the storage unit and the computing unit = {F ij |i, j = 1, 2, ..., N}, where F ij Represents the total frequency of data transferred from the storage unit to the computing unit by task i being accessed by task j:

[0055]

[0056] Among them, R ij (t, s) is the resource matching matrix between task i and task j at time t and spatial position s, and Δt=t2-t1 is the length of the time window.

[0057] S34, defining each cache unit c based on the total frequency and the state of the cache unit k The cache strategy P k ={p ij，k |i, j = 1, 2, ..., N}, where p ij,k Indicates that the data of task i is in cache unit c k The priority of the cache:

[0058]

[0059] Among them, α2 and β2 are weight coefficients, indicating the importance of data access frequency and cache frequency. Cache unit c k Available storage capacity;

[0060] S35, based on cache strategy P k Dynamically adjust cache unit c k The storage location of the data in the selection p ij,k The largest task data is cached first, and p is removed at the same time. ij,k Minimum data to free up cache space;

[0061] S36. Each cache unit c on the transmission path kIn the process, the data access frequency matrix A is updated in real time through the feedback adjustment mechanism k and available storage capacity And recalculate the cache strategy P according to the task load change k .

[0062] Optionally, the S4 includes:

[0063] S41, caching strategy P based on intelligent cache unit k Analyze the access frequency and transmission requirements of different tasks on the data transmission path, and define the task access frequency matrix T on the data transmission path = {T ij |i, j = 1, 2, ..., N}, where T ij Represents the total frequency of task i’s data being accessed by task j on the path:

[0064]

[0065] Among them, A ij,k The data of task i is in cache unit c k The frequency of being accessed by task j, Δd k Cache unit c k The distance to the target cell;

[0066] S42, based on the task access frequency matrix T and the state of the cache unit, define a dynamic cache strategy matrix P′={p′ ij，k |i, j = 1, 2, ..., N; k = 1, 2, ..., K}, where p′ ij,k Indicates that the data of task i is cached in unit c on the path k Adjustment priority:

[0067]

[0068] Among them, α3, β3, γ2, and δ2 are weight coefficients, which respectively represent the importance of access frequency, cache frequency, available storage capacity, and transmission path distance, d k Cache unit c k The distance to the mission target unit;

[0069] S43, according to the dynamic cache strategy matrix P', the cache position of the high-frequency access data on the data transmission path is adjusted, and the high priority p' is ij，k The task data of p′ is cached to the cache unit closer to the target unit, and the low priority p′ is cached to the cache unit closer to the target unit. ij，k The data is gradually removed to release the cache unit space;

[0070] S44, the adjusted cache unit c kThe state is updated in real time, and the access frequency matrix A of the cache unit is recalculated. k and available storage capacity And dynamically update the cache strategy P' according to the changes of the task access frequency matrix T to adapt to the new task load and path requirements;

[0071] S45. In the process of dynamic cache strategy adjustment, the time-space deep collaborative game model is used to generate the global optimal path cache solution to minimize the distance of the data transmission path and the cache energy consumption.

[0072] Optionally, the S8 includes:

[0073] S81. Based on the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, define the task priority matrix P task = {P i |i=1,2,...,N}, where P i Indicates the priority of task i:

[0074]

[0075] in, The computation, storage and transmission resource requirements of task i are respectively: are the currently available computing, storage and transmission resources, respectively, c ,ω m ,ω b is the task priority weight coefficient;

[0076] S82, based on the task priority matrix P task and the state matrix of the cache unit to construct the resource competition matrix R comp = {R ij |i, j = 1, 2, ..., N}, where R ij Indicates the degree of competition between task i and task j in resource allocation:

[0077]

[0078] Among them, T ij A represents the total frequency of task i’s data being accessed by task j on the path. ij，k is the cache unit T ij The data access frequency of task i and task j, α4, β4, γ3 are the weight coefficients for calculating resource contention;

[0079] S83, based on resource competition matrix R comp And the spatiotemporal deep collaborative game model, define the global resource allocation utility function U global :

[0080]

[0081] Among them, E i is the total energy consumption of task i:

[0082]

[0083] Among them, δ2 and λ3 are weight coefficients of the utility function, which are used to balance resource competition and energy consumption;

[0084] S84, by optimizing the global resource allocation utility function U global Generate the global optimal resource allocation strategy S opt , including task priority, cache unit location, and path selection:

[0085] S opt = arg maxU global ;

[0086] Among them, the constraints include task resource demand constraints:

[0087]

[0088]

[0089]

[0090] Cache capacity constraints:

[0091]

[0092] S85, the generated global optimal resource allocation strategy S opt Applied to AI servers, it dynamically adjusts task priorities, cache unit allocation, and data transmission paths to optimize the overall energy efficiency of the AI ​​server to the optimal state.

[0093] The beneficial effects of the present invention are:

[0094] (1) The present invention introduces a spatiotemporal deep collaborative game model for the first time, treating the storage unit, computing unit, and cache unit in the AI ​​server as multiple intelligent agents, and optimizing the global resource utilization efficiency under multi-task loads by dynamically adjusting the resource allocation strategy. Traditional resource allocation methods usually lack a comprehensive consideration of the resource competition status between tasks. The present invention realizes the collaborative optimization of computing resources, storage resources, and data transmission resources by constructing a task priority matrix and a resource competition matrix combined with an intelligent caching strategy, so that resource configuration dynamically adapts to changes in task requirements, effectively reducing resource waste and performance bottlenecks caused by load conflicts.

[0095] (2) The present invention designs a dynamic caching strategy based on access frequency and path priority to meet the frequent data transmission needs in AI servers. By defining the task access frequency matrix and the cache strategy matrix, intelligent cache adjustment is achieved on the bidirectional data transmission path. Different from the static caching scheme in the prior art, the present invention dynamically adjusts the cache position according to the high-frequency and low-frequency data access modes of the task, and preferentially caches high-frequency data near the target unit, effectively reducing the data transmission distance and energy consumption. Experiments show that the data transmission efficiency is improved by more than 30% and the transmission energy consumption is reduced by more than 15% by the method of the present invention.

[0096] (3) The present invention realizes the global optimal resource configuration strategy in task priority, cache unit allocation and path selection by optimizing the global resource allocation utility function. In traditional methods, global resource optimization is often limited by the efficiency of single task resource allocation, and it is difficult to take into account the overall performance of the system. The present invention combines the multi-dimensional characteristics of task priority weight, cache unit capacity and inter-task transmission requirements to generate a dynamic adjustment plan to achieve a dynamic balance between resource competition and energy efficiency. Experimental results show that after adopting the optimization strategy of the present invention, the overall energy efficiency of the AI ​​server is improved by 25%, and the response speed in a complex multi-task environment is improved by 20%. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0098] Figure 1 This is a flow chart of a method for optimizing the energy efficiency of a storage, computing and training AI server based on deep spatiotemporal collaborative game proposed by the present invention;

[0099] Figure 2 This is a schematic diagram of the global resource allocation optimization framework based on the space-time deep collaborative game model in the storage, computing and training AI server energy efficiency optimization method based on space-time deep collaborative game proposed in the present invention. DETAILED DESCRIPTION

[0100] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0101] refer to Figure 1-Figure 2 , a method for optimizing the energy efficiency of storage and computing training AI servers based on deep spatiotemporal collaborative game, including the following:

[0102] S1. Obtain a multi-task load information set of the AI ​​server, including resource requirements of parallel tasks, data transmission requirements, and computing resource usage;

[0103] S2. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set. The spatiotemporal deep collaborative game model includes multiple agents for representing the data flow between the storage unit and the computing unit in the AI ​​server. Each agent corresponds to a different storage unit or computing unit to represent its own data storage and transmission requirements.

[0104] S3. An intelligent cache unit is set on the path where data is transmitted from the storage unit to the computing unit and returned from the computing unit to the storage unit. The intelligent cache unit analyzes the data access frequency between the storage unit and the computing unit based on the spatiotemporal depth collaborative game model, dynamically adjusts the storage location and cache strategy of the cached data, and realizes intelligent data caching on the bidirectional transmission path;

[0105] S4. Analyze the access frequency and transmission requirements of different tasks on the data transmission path based on the spatiotemporal deep collaborative game model, and generate a dynamic cache strategy, which includes the intelligent cache unit adjusting the cache position of high-frequency access data and low-frequency access data on different transmission paths;

[0106] S5. According to the dynamic cache strategy, high-frequency access data is preferentially cached in the cache unit close to the computing unit, and low-frequency access data is pre-stored in the cache unit close to the storage unit to optimize the bidirectional data transmission path;

[0107] S6. When the task load changes, the new task load information and its resource demand changes for storage units and computing units are analyzed in real time through the spatiotemporal deep collaborative game model, and the cache strategy of the intelligent cache unit is updated based on the analysis results, and the cache location and path of high-frequency access data are reallocated to adapt the cache strategy to the optimal data flow under the current load conditions;

[0108] S7. Automatically release redundant cache data in the intelligent cache unit after executing the cache strategy update, and monitor the load of the data transmission path in real time through the feedback adjustment mechanism of the spatiotemporal deep collaborative game model to optimize the energy efficiency performance of the data transmission path from the storage unit to the computing unit and the return path;

[0109] S8. Under the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, the resource competition status between different task loads in the AI ​​server is further analyzed, and the global optimal resource allocation strategy is generated in terms of task priority, cache unit location and path selection to improve the overall energy efficiency of the AI ​​server to the optimal state.

[0110] In this implementation, S1 includes:

[0111] S11. Obtain resource requirement information of each parallel task in the AI ​​server. The resource requirement information includes the computing resource requirement C of each task.i , Storage resource requirements M i And data transmission requirements D i , where i is the task identifier, i=1, 2, ..., N, and N is the total number of current parallel tasks;

[0112] S12. Constructing multi-task load information set:

[0113] L={(C i , M i , D i )|i=1,2,...,N}.

[0114] In this implementation, S2 includes:

[0115] S21. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set L and construct a multi-agent set A = {a j |j=1,2,...,K}, where K is the total number of agents, and each agent a j Corresponds to a storage unit or computing unit in the AI ​​server, used to represent its data storage and transmission requirements at time t and spatial location s;

[0116] S22. For each agent a j Define the strategy space S j , its strategies j (t,s)∈S j Represents agent a j Resource allocation scheme at time t and spatial location s:

[0117] Computational resource allocationc j (t, s) represents agent a j The amount of computing resources allocated at time t and location s;

[0118] Storage resource allocation j (t, s) represents agent a j The amount of storage resources allocated at time t and location s;

[0119] Data transmission bandwidth allocationb j (t, s) represents agent a j The data transmission bandwidth allocated at time t and location s;

[0120] S23. Define the resource requirement function for each task i:

[0121] D i (t, s) = (C i (t,s),M i (t,s),B i (t, s));

[0122] Among them, C i (t, s) represents the computing resource requirements of task i at time t and location s, M i (t, s) represents the storage resource requirement of task i at time t and location s, B i (t, s) represents the data transmission bandwidth requirement of task i at time t and location s;

[0123] S24, establish a resource matching matrix R between tasks and agents to represent the resource allocation relationship between tasks and agents in the spatiotemporal dimension ij (t, s):

[0124]

[0125] S25. Define agent a j The utility function U j (t, s), used to measure its resource utilization efficiency:

[0126]

[0127] Among them, N is the total number of tasks, α1, β1, γ1 are weight coefficients, indicating the importance of computing resources, storage resources and transmission bandwidth, δ1 is the energy penalty coefficient, which is used to balance the relationship between resource utilization efficiency and energy consumption, E j (t, s) is the agent a j Energy consumption at time t and location s:

[0128] E j (t, s) = η c ·c j (t, s)+η m ·m j (t, s)+η b b j (t, s);

[0129] Among them, η c、 η m , η b are the energy consumption coefficients of computing resources, storage resources, and transmission bandwidth respectively;

[0130] S26. Determine the optimal strategy for each agent by solving the optimization problem To maximize the total system utility:

[0131]

[0132] Among them, the constraints are task resource requirements constraints:

[0133]

[0134]

[0135]

[0136] Agent resource capacity constraints:

[0137]

[0138]

[0139]

[0140]

[0141] in, They are respectively j The maximum computing resources, storage resources and data transmission bandwidth capacity, s j (t, s) is the agent a j The strategy includes the resource allocation scheme at time t and location s.

[0142] In this implementation, S3 includes:

[0143] S31, set a smart cache unit set C = {c k |k=1,2,...,K}, where K is the total number of cache units, and the smart cache unit c k An intermediate node located in the transmission path, used to temporarily store data from the storage unit to the computing unit and from the computing unit back to the storage unit;

[0144] S32. Define each cache unit c k The initial state variables are:

[0145]

[0146] in, Represents cache unit c k The currently available free storage capacity, A k Represents cache unit c k The data access frequency matrix of the internal cache, the element A of the matrix ij,k Indicates that the data of task i is in node c k The frequency of being visited by task j;

[0147] S33, based on the spatiotemporal depth collaborative game model, calculate the access frequency matrix F between the storage unit and the computing unit = {F ij |i, j = 1, 2, ..., N}, where F ijRepresents the total frequency of data transferred from the storage unit to the computing unit by task i being accessed by task j:

[0148]

[0149] Among them, R ij (t, s) is the resource matching matrix between task i and task j at time t and spatial position s, and Δt=t2-t1 is the length of the time window.

[0150] S34, defining each cache unit c based on the total frequency and the state of the cache unit k The cache strategy P k ={p ij，k |i, j = 1, 2, ..., N}, where p ij,k Indicates that the data of task i is in cache unit c k The priority of the cache:

[0151]

[0152] Among them, α2 and β2 are weight coefficients, indicating the importance of data access frequency and cache frequency. Cache unit c k Available storage capacity;

[0153] S35, based on cache strategy P k Dynamically adjust cache unit c k The storage location of the data in the selection p ij,k The largest task data is cached first, and p is removed at the same time. ij,k Minimum data to free up cache space;

[0154] S36. Each cache unit c on the transmission path k In the process, the data access frequency matrix A is updated in real time through the feedback adjustment mechanism k and available storage capacity And recalculate the cache strategy P according to the task load change k .

[0155] In this implementation, S4 includes:

[0156] S41, caching strategy P based on intelligent cache unit k Analyze the access frequency and transmission requirements of different tasks on the data transmission path, and define the task access frequency matrix T on the data transmission path = {T ij |i, j = 1, 2, ..., N}, where T ij Represents the total frequency of task i’s data being accessed by task j on the path:

[0157]

[0158] Among them, A ij,k The data of task i is in cache unit c k The frequency of being accessed by task j, Δd k Cache unit c k The distance to the target cell;

[0159] S42, based on the task access frequency matrix T and the state of the cache unit, define a dynamic cache strategy matrix P′={p′ ij，k |i, j = 1, 2, ..., N; k = 1, 2, ..., K}, where p′ ij，k Indicates that the data of task i is cached in unit c on the path k Adjustment priority:

[0160]

[0161] Among them, α3, β3, γ2, and δ2 are weight coefficients, which respectively represent the importance of access frequency, cache frequency, available storage capacity, and transmission path distance, d k Cache unit c k The distance to the mission target unit;

[0162] S43, according to the dynamic cache strategy matrix P', the cache position of the high-frequency access data on the data transmission path is adjusted, and the high priority p' is ij，k The task data of p′ is cached to the cache unit closer to the target unit, and the low priority p′ is cached to the cache unit closer to the target unit. ij，k The data is gradually removed to release the cache unit space;

[0163] S44, the adjusted cache unit c k The state is updated in real time, and the access frequency matrix A of the cache unit is recalculated. k and available storage capacity And dynamically update the cache strategy P' according to the changes of the task access frequency matrix T to adapt to the new task load and path requirements;

[0164] S45. In the process of dynamic cache strategy adjustment, the time-space deep collaborative game model is used to generate the global optimal path cache solution to minimize the distance of the data transmission path and the cache energy consumption.

[0165] In this implementation, S8 includes:

[0166] S81. Based on the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, define the task priority matrix P task = {P i |i=1,2,...,N}, where P i Indicates the priority of task i:

[0167]

[0168] in, The computation, storage and transmission resource requirements of task i are respectively: are the currently available computing, storage and transmission resources, respectively, c ,ω m ,ω b is the task priority weight coefficient;

[0169] S82, based on the task priority matrix P task and the state matrix of the cache unit to construct the resource competition matrix R comp = {R ij |i, j = 1, 2, ..., N}, where R ij Indicates the degree of competition between task i and task j in resource allocation:

[0170]

[0171] Among them, T ij A represents the total frequency of task i’s data being accessed by task j on the path. ij，k is the cache unit T ij The data access frequency of task i and task j, α4, β4, γ3 are the weight coefficients for calculating resource contention;

[0172] S83, based on resource competition matrix R comp And the spatiotemporal deep collaborative game model, define the global resource allocation utility function U global :

[0173]

[0174] Among them, E i is the total energy consumption of task i:

[0175]

[0176] Among them, δ2 and λ3 are weight coefficients of the utility function, which are used to balance resource competition and energy consumption;

[0177] S84, by optimizing the global resource allocation utility function U global Generate the global optimal resource allocation strategy S opt , including task priority, cache unit location, and path selection:

[0178] S opt = arg maxU global ;

[0179] Among them, the constraints include task resource demand constraints:

[0180]

[0181]

[0182]

[0183] Cache capacity constraints:

[0184]

[0185] S85, the generated global optimal resource allocation strategy S opt Applied to AI servers, it dynamically adjusts task priorities, cache unit allocation, and data transmission paths to optimize the overall energy efficiency of the AI ​​server to the optimal state.

[0186] Embodiment 1:

[0187] In the data center of a large financial institution, AI servers are widely used for financial data prediction, risk assessment, and customer behavior analysis tasks. Since these tasks usually have the characteristics of high concurrency and dynamic load, the server needs to process requests from multiple departments at the same time, including real-time data processing and historical data analysis. In this scenario, data is frequently transmitted between storage units and computing units, resource competition is fierce, and system energy efficiency and resource utilization need to be optimized urgently.

[0188] In one application, the risk assessment department needs to process more than 10 million transaction records per day during peak trading hours. The data needs to be retrieved from the distributed storage unit to the computing unit for analysis. At the same time, the customer behavior analysis task requires real-time processing of large-scale user behavior logs, which have high data transmission and storage requirements. In this complex task environment, the traditional static resource allocation strategy is prone to resource waste, task delays and increased energy consumption.

[0189] This embodiment optimizes the above scenario by deploying a storage, computing and training AI server energy efficiency optimization method based on a spatiotemporal deep collaborative game model.

[0190] First, the AI ​​server in the data center obtains the resource requirement information of multiple tasks, including the computing resource requirements of real-time transaction data, the storage resource requirements of historical analysis data, and the transmission bandwidth requirements of behavior log data. The requirement information of each task is sorted through the task priority matrix, where transaction risk assessment is set as a high-priority task, while customer behavior analysis and historical data analysis tasks have lower priorities.

[0191] Next, the spatiotemporal deep collaborative game model initializes multiple intelligent agents, corresponding to different computing units, storage units and cache units respectively. During peak hours, the system recognizes that transaction data transmission needs are frequent, and the cache strategy matrix prioritizes caching high-frequency access transaction data to the cache unit closest to the computing unit. For customer behavior log data, due to its low access frequency, the cache strategy prioritizes storing it in the cache unit far away from the computing unit.

[0192] Subsequently, the system analyzes the resource competition status and adjusts the task resource configuration in real time by optimizing the global resource allocation utility function. When the bandwidth demand of the transaction risk assessment task surges, the system dynamically adjusts the cache location and data transmission path to release some resources for non-urgent tasks so that high-priority tasks can complete the analysis quickly. At the same time, historical data analysis tasks are scheduled to be performed during the idle time of the computing unit during non-peak hours, further improving resource utilization.

[0193] In order to verify the effectiveness of the present invention, two comparison methods are set up in the experiment: one is to use the traditional static resource allocation method, and the other is to use the optimization method of the present invention.

[0194] Experimental conditions: Location: a large financial data center, Time: September 2024, Dataset: 50 million daily transaction data, 10 million behavior logs, 1TB of historical analysis data, Test duration: 30 consecutive days:

[0195] Table 1 Data comparing the traditional method with the method of the present invention in the experiment

[0196] index Traditional methods Method of the present invention Improvement ratio Average task completion time (seconds) 15.6 10.3 34％ Data transmission delay during peak hours (seconds) 0.8 0.3 62.5％ System average energy consumption (kWh / day) 3500 2800 20％ High-frequency data cache hit rate 68％ 92％ 35.3％ Average task resource utilization 72％ 88％ 22.2％

[0197] As can be seen from Table 1, in the 30-day test, the optimization method of the present invention is significantly better than the traditional method in all indicators. The average task completion time is reduced from 15.6 seconds to 10.3 seconds, an increase of 34%. The data transmission delay during peak hours is reduced from 0.8 seconds to 0.3 seconds, an increase of 62.5%. In terms of energy efficiency, the average energy consumption of the system is reduced by 20%, saving about 700 kWh per day. At the same time, the high-frequency data cache hit rate is significantly improved from 68% of the traditional method to 92%, further reducing the time and energy consumption of data transmission. The improvement in resource utilization also reflects the optimization effect of the present invention on the resource competition problem in a multi-tasking environment.

[0198] Through simulation tests on a large financial data center, this Example 1 proves the efficiency and feasibility of the method of the present invention in a dynamic load environment. The dynamic resource allocation strategy and intelligent cache optimization strategy of the present invention show significant advantages in improving system energy efficiency, shortening task completion time and reducing data transmission delay, providing technical support for the efficient operation of AI servers in complex task environments.

[0199] The present invention introduces a deep spatiotemporal collaborative game model for the first time, treating the storage unit, computing unit and cache unit in the AI ​​server as multiple intelligent agents, and optimizing the global resource utilization efficiency under multi-task loads by dynamically adjusting the resource allocation strategy. Traditional resource allocation methods usually lack a comprehensive consideration of the resource competition status between tasks. The present invention realizes the collaborative optimization of computing resources, storage resources and data transmission resources by constructing a task priority matrix and a resource competition matrix combined with an intelligent caching strategy, so that resource configuration dynamically adapts to changes in task requirements, effectively reducing resource waste and performance bottlenecks caused by load conflicts.

[0200] The present invention designs a dynamic caching strategy based on access frequency and path priority in response to the frequent data transmission needs in AI servers. By defining the task access frequency matrix and the cache strategy matrix, intelligent cache adjustment is realized on the bidirectional transmission path of data. Different from the static caching scheme in the prior art, the present invention dynamically adjusts the cache position according to the high-frequency and low-frequency data access modes of the task, and preferentially caches high-frequency data near the target unit, effectively reducing the data transmission distance and energy consumption. Experiments show that the data transmission efficiency is improved by more than 30% and the transmission energy consumption is reduced by more than 15% by the method of the present invention.

[0201] The present invention realizes the global optimal resource configuration strategy in task priority, cache unit allocation and path selection by optimizing the global resource allocation utility function. In traditional methods, global resource optimization is often limited by the efficiency of single task resource allocation, and it is difficult to take into account the overall performance of the system. The present invention combines the multi-dimensional characteristics of task priority weights, cache unit capacity and inter-task transmission requirements to generate a dynamic adjustment plan to achieve a dynamic balance between resource competition and energy efficiency. Experimental results show that after adopting the optimization strategy of the present invention, the overall energy efficiency of the AI ​​server is improved by 25%, and the response speed in a complex multi-task environment is improved by 20%.

[0202] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for optimizing the energy efficiency of AI servers based on deep spatial-temporal collaborative game, characterized in that: The steps include: S1. Obtain a multi-task load information set of the AI ​​server; S2. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set, the spatiotemporal deep collaborative game model includes multiple agents for representing the data flow between the storage unit and the computing unit in the AI ​​server, each agent corresponding to a different storage unit or computing unit; S3. An intelligent cache unit is set on the path where data is transmitted from the storage unit to the computing unit and returned from the computing unit to the storage unit. The intelligent cache unit analyzes the data access frequency between the storage unit and the computing unit based on the spatiotemporal depth collaborative game model and dynamically adjusts the storage location and cache strategy of the cached data; S4. Analyze the access frequency and transmission requirements of different tasks on the data transmission path based on the spatiotemporal deep collaborative game model, and generate a dynamic cache strategy; S5. Preferentially cache high-frequency access data in a cache unit close to the computing unit according to a dynamic cache strategy, and pre-store low-frequency access data in a cache unit close to the storage unit; S6. When the task load changes, the new task load information and its resource demand changes for the storage unit and computing unit are analyzed in real time through the spatiotemporal deep collaborative game model, and the cache strategy of the intelligent cache unit is updated based on the analysis results to reallocate the cache location and path of high-frequency access data; S7. Automatically release redundant cache data in the intelligent cache unit after executing the cache strategy update, and monitor the load of the data transmission path in real time through the feedback adjustment mechanism of the spatiotemporal depth collaborative game model; S8. Further analyze the resource competition status between different task loads in the AI ​​server under the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, and generate a global optimal resource allocation strategy based on task priority, cache unit location, and path selection.

2. According to claim 1, a method for optimizing the energy efficiency of storage, computing and training AI servers based on time-space deep collaborative game, characterized in that: The S1 includes: S11, obtaining resource requirement information of each parallel task in the AI ​​server, wherein the resource requirement information includes the computing resource requirement C of each task i , Storage resource requirements M i And data transmission requirements D i , where i is the task identifier, i = 1, 2, ..., N, and N is the total number of current parallel tasks; S12. Constructing multi-task load information set: L={(C i ,M i ,D i )∣i=1,2,...,N}。 3. According to claim 1, a method for optimizing the energy efficiency of storage, computing and training AI servers based on time-space deep collaborative game, characterized in that: The S2 includes: S21. Initialize the spatiotemporal deep collaborative game model based on the multi-task load information set L and construct a multi-agent set A = {a j |j=1,2,...,K}, where K is the total number of agents, and each agent a j Corresponds to a storage unit or computing unit in the AI ​​server, used to represent its data storage and transmission requirements at time t and spatial location s; S22. For each agent a j Define the strategy space S j , its strategies j (t,s)∈S j Represents agent a j Resource allocation scheme at time t and spatial location s: Computational resource allocationc j (t,s) represents agent a j The amount of computing resources allocated at time t and location s; Storage resource allocation j (t,s) represents agent a j The amount of storage resources allocated at time t and location s; Data transmission bandwidth allocationb j (t,s) represents agent a j The data transmission bandwidth allocated at time t and location s; S23. Define the resource requirement function for each task i: D i (t,s)=(C i (t,s),M i (t,s),B i (t,s)); Among them, C i (t,s) represents the computing resource requirements of task i at time t and location s, M i (t,s) represents the storage resource requirement of task i at time t and location s, B i (t, s) represents the data transmission bandwidth requirement of task i at time t and location s; S24, establish a resource matching matrix R between tasks and agents to represent the resource allocation relationship between tasks and agents in the spatiotemporal dimension ij (t,s): S25. Define agent a j The utility function U j (t,s), used to measure its resource utilization efficiency: Among them, N is the total number of tasks, α1, β1, γ1 are weight coefficients, indicating the importance of computing resources, storage resources and transmission bandwidth, δ1 is the energy penalty coefficient, which is used to balance the relationship between resource utilization efficiency and energy consumption, E j (t,s) is agent a j Energy consumption at time t and location s: E j (t,s)=η c ·c j (t,s)+η m ·m j (t,s)+η b ·b j (t,s); Among them, η c , η m , η b are the energy consumption coefficients of computing resources, storage resources, and transmission bandwidth respectively; S26. Determine the optimal strategy for each agent by solving the optimization problem To maximize the total system utility: Among them, the constraints are task resource requirements constraints: Agent resource capacity constraints: in, They are respectively j The maximum computing resources, storage resources and data transmission bandwidth capacity, s j (t,s) is agent a j The strategy includes the resource allocation scheme at time t and location s.

4. According to claim 1, a method for optimizing the energy efficiency of storage, computing and training AI servers based on time-space deep collaborative game, characterized in that: The S3 includes: S31, set a smart cache unit set C = {c k |k=1,2,...,K}, where K is the total number of cache units, and the number of smart cache units c k An intermediate node located in the transmission path, used to temporarily store data from the storage unit to the computing unit and from the computing unit back to the storage unit; S32. Define each cache unit c k The initial state variables are: in, Represents cache unit c k The currently available free storage capacity, A k Represents cache unit c k The data access frequency matrix of the internal cache, the element A of the matrix ij,k Indicates that the data of task i is in node c k The frequency of being visited by task j; S33, based on the spatiotemporal depth collaborative game model, calculate the access frequency matrix F between the storage unit and the computing unit = {F ij |i,j=1,2,...,N}, where F ij Represents the total frequency of data transferred from the storage unit to the computing unit by task i being accessed by task j: Among them, R ij (t, s) is the resource matching matrix between task i and task j at time t and spatial position s, and Δt = t2-t1 is the length of the time window. S34, defining each cache unit c based on the total frequency and the state of the cache unit k The cache strategy P k ={p ij,k |i,j=1,2,...,N}, where p ij,k Indicates that the data of task i is in cache unit c k The priority of the cache: Among them, α2 and β2 are weight coefficients, indicating the importance of data access frequency and cache frequency. Cache unit c k Available storage capacity; S35, based on cache strategy P k Dynamically adjust cache unit c k The storage location of the data in the selection p ij,k The largest task data is cached first, and p is removed at the same time. ij,k Minimum data to free up cache space; S36. Each cache unit c on the transmission path k In the process, the data access frequency matrix A is updated in real time through the feedback adjustment mechanism k and available storage capacity And recalculate the cache strategy P according to the task load change k .

5. According to claim 1, a method for optimizing the energy efficiency of storage, computing and training AI servers based on time-space deep collaborative game, characterized in that: The S4 includes: S41, caching strategy P based on intelligent cache unit k Analyze the access frequency and transmission requirements of different tasks on the data transmission path, and define the task access frequency matrix T on the data transmission path = {T ij |i,j=1,2,...,N}, where T ij Represents the total frequency of task i’s data being accessed by task j on the path: Among them, A ij,k The data of task i is in cache unit c k The frequency of being accessed by task j, Δd k Cache unit c k The distance to the target cell; S42: Define a dynamic cache strategy matrix P′={p′ based on the task access frequency matrix T and the state of the cache unit. ij,k |i,j=1,2,...,N;k=1,2,...,K}, where p′ ij,k Indicates that the data of task i is cached in unit c on the path k Adjustment priority: Among them, α3, β3, γ2, and δ2 are weight coefficients, which respectively represent the importance of access frequency, cache frequency, available storage capacity, and transmission path distance, d k Cache unit c k The distance to the mission target unit; S43, according to the dynamic cache strategy matrix P', the cache position of the high-frequency access data on the data transmission path is adjusted, and the high priority p' is ij,k The task data of p′ is cached to the cache unit closer to the target unit, and the low priority p′ is cached to the cache unit closer to the target unit. ij,k The data is gradually removed to release the cache unit space; S44, the adjusted cache unit c k The state is updated in real time, and the access frequency matrix A of the cache unit is recalculated. k and available storage capacity And dynamically update the cache strategy P' according to the changes of the task access frequency matrix T to adapt to the new task load and path requirements; S45. In the process of dynamic cache strategy adjustment, the time-space deep collaborative game model is used to generate the global optimal path cache solution to minimize the distance of the data transmission path and the cache energy consumption.

6. According to claim 1, a method for optimizing the energy efficiency of storage, computing and training AI servers based on time-space deep collaborative game, characterized in that: The S8 includes: S81. Based on the feedback adjustment mechanism of the spatiotemporal deep collaborative game model, define the task priority matrix P task = {P i |i=1,2,...,N}, where P i Indicates the priority of task i: in, The computation, storage and transmission resource requirements of task i are respectively: are the currently available computing, storage and transmission resources, respectively, c ,ω m ,ω b is the task priority weight coefficient; S82, based on the task priority matrix P task and the state matrix of the cache unit to construct the resource competition matrix R comp = {R ij |i,j=1,2,...,N}, where R ij Indicates the degree of competition between task i and task j in resource allocation: Among them, T ij A represents the total frequency of task i’s data being accessed by task j on the path. ij,k is the cache unit T ij The data access frequency of task i and task j, α4, β4, γ3 are the weight coefficients for calculating resource contention; S83, based on resource competition matrix R comp And the spatiotemporal deep collaborative game model, define the global resource allocation utility function U global : Among them, E i is the total energy consumption of task i: Among them, δ2 and λ3 are weight coefficients of the utility function, which are used to balance resource competition and energy consumption; S84, by optimizing the global resource allocation utility function U global Generate the global optimal resource allocation strategy S opt , including task priority, cache unit location, and path selection: S opt =argmaxU global ; Among them, the constraints include task resource demand constraints: Cache capacity constraints: S85, the generated global optimal resource allocation strategy S opt Applied to AI servers, it dynamically adjusts task priorities, cache unit allocation, and data transmission paths to optimize the overall energy efficiency of the AI ​​server to the optimal state.

Images