Storage, calculation and training AI server load fluctuation self-balancing method based on recursive game network

Abstract

The invention discloses a storage, calculation and training AI server load fluctuation self-balancing method based on a recursive game network. The method comprises the following steps: S1, recording real-time load data of each task; s2, constructing a macroscopic hierarchical recursive game network model based on the overall resource pool of the AI server; s3, generating an overall resource allocation strategy through multi-agent competition calculation of the macroscopic hierarchical recursive game network model; s4, constructing a micro-hierarchy recursive game network model based on the resource allocation strategy generated by the macro-hierarchy; s5, sub-task resources are optimally allocated; s6, issuing the adjusted resource quota to the microscopic hierarchical recursive game network model in real time; s7, carrying out resource pre-allocation on the upcoming high-load task; and S8, enabling resource allocation to gradually tend to an optimal state. According to the method, the resource utilization efficiency, the system response capability and the task completion stability of the AI server in a multi-task load fluctuation scene are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AI servers, and in particular to a method for self-balancing load fluctuations of a storage, computing and training AI server based on a recursive game network. Background Art

[0002] With the rapid development of artificial intelligence technology and high-performance computing, AI servers have become an important infrastructure to support large-scale data processing and deep learning model training. In practical applications, AI servers need to handle multi-task computing, storage, and training requirements simultaneously, which involves a large amount of resource allocation and dynamic regulation. However, achieving efficient and stable resource management under complex and changing load environments is still a problem that needs to be solved urgently.

[0003] At present, traditional server load balancing technology mainly relies on static or simple dynamic resource allocation algorithms. Traditional methods usually allocate the server's computing resources, storage resources and training resources to different task modules according to preset proportions based on fixed load balancing rules. However, the fixed rule allocation method shows significant limitations when facing actual scenarios with multiple tasks and high load fluctuations. When the resource demand of certain tasks suddenly increases, traditional methods are often unable to adjust resource allocation in time, resulting in performance degradation of high-priority tasks due to insufficient resources.

[0004] In recent years, some dynamic load balancing algorithms have begun to be applied to AI servers. Dynamic load balancing algorithms dynamically adjust server resource allocation by monitoring task load changes in real time. However, in actual applications, dynamic load balancing algorithms usually only focus on adjusting the current load and lack the ability to predict load fluctuation trends, resulting in slow resource regulation when facing load peaks. On the other hand, most of the existing dynamic load balancing algorithms cannot achieve joint regulation of computing resources, storage resources, and training resources, which can easily lead to low system resource utilization or bottlenecks.

[0005] In summary, the existing technologies have obvious deficiencies in load prediction capabilities, multi-resource joint regulation, and resource allocation efficiency, and are unable to meet the efficient and stable operation requirements of modern AI servers in complex scenarios. Therefore, there is an urgent need for a smarter and more efficient resource regulation method to cope with the resource management challenges in AI server multi-task load fluctuation scenarios. Summary of the invention

[0006] One purpose of the present invention is to propose a method for self-balancing the load fluctuation of a storage, computing and training AI server based on a recursive game network. The present invention significantly improves the resource utilization efficiency, system responsiveness and task completion stability of the AI ​​server in a multi-task load fluctuation scenario.

[0007] According to an embodiment of the present invention, a method for self-balancing load fluctuations of a storage and computing training AI server based on a recursive game network includes the following steps:

[0008] S1. Monitor each computing task in the AI ​​server, collect the computing resource requirements, data storage resource requirements, and training resource requirements of each computing task in real time, and record the real-time load data of each task;

[0009] S2. Construct a macro-level recursive game network model based on the overall resource pool of the AI ​​server, take the different AI models carried by the AI ​​server as game agents, and allocate and control the total amount of computing resources, storage resources, and training resources of the AI ​​server;

[0010] S3. Generate an overall resource allocation strategy through multi-agent competitive calculation of a macro-level recursive game network model, determine the initial computing resource allocation, storage resource allocation, and training resource allocation for each AI model based on the task priority and resource requirements of each AI model in the server, and dynamically respond to the prediction of load fluctuations;

[0011] S4. Based on the resource allocation strategy generated at the macro level, a micro-level recursive game network model is constructed. The specific computing task module, data storage module, and training task module in the AI ​​server are defined as micro-agents. Under the resource quota provided by the macro level, each agent plays a game based on the resource requirements and load fluctuations of the task, and achieves real-time load regulation and resource allocation for each subtask through competitive behavior.

[0012] S5. In the micro-level recursive game network model, each micro-agent analyzes the load of the current subtask in real time, calculates the resource occupancy rate and idle rate, and inputs the feedback results into the micro-level recursive game network model. Each micro-agent dynamically adjusts the resource usage ratio according to the feedback resource usage and task load changes through an adaptive control mechanism;

[0013] S6. Combining the game results of the macro-level recursive game network model and the micro-level recursive game network model, when it is detected that a subtask has insufficient resources or the load exceeds the current quota, the micro-level recursive game network model submits a resource demand adjustment request to the macro-level recursive game network model, and the macro-level recursive game network model reallocates the overall resources of the AI ​​model and sends the adjusted resource quota to the micro-level recursive game network model in real time;

[0014] S7. Based on the historical load data and the real-time load data in the macro-level recursive game network model and the micro-level recursive game network model, the future load fluctuation trend is predicted through a recursive feedback model to generate a trend estimate of resource demand;

[0015] S8. During the load balancing operation of the AI ​​server, dynamically adjust the resource control strategies of the macro-level recursive game network model and the micro-level recursive game network model, gradually optimize the resource decisions of the game agents, and make the resource allocation gradually tend to the optimal state.

[0016] Optionally, the S1 includes the following steps:

[0017] S11. Monitor each computing task in the AI ​​server in real time, collect computing resource requirements, data storage resource requirements, and training resource requirements of the computing task, and the real-time load data of the computing task is expressed as L i (t), where i is the unique identifier of the task and t is the time point;

[0018] S12. Record the real-time computing resource requirements, data storage resource requirements, and training resource requirements of each computing task respectively, and obtain the comprehensive load data vector D of each task. i (t):

[0019] D i (t) = {C i (t),M i (t),T i (t)};

[0020] Among them, C i (t) represents the amount of computing resources required by task i at time t, in floating-point operations, M i (t) represents the amount of data storage resources required by task i at time t, in bytes, T i (t) represents the amount of training resources required for task i at time t, in training steps;

[0021] S13, aggregating the real-time load data of multiple tasks at time point t to form a time series task load data set D(t);

[0022] S14. Record the load change trend of each task and calculate the load change rate Δ of task i i (t):

[0023]

[0024] Among them, Δt is the time interval of load collection;

[0025] S15, input the real-time collected task load data, comprehensive load data vector and load change rate into the dynamic load data storage module to form a task real-time load data set

[0026]

[0027] Among them, Δ(t)={Δ1(t),Δ2(t),…,Δ n (t)} is the set of collected time points.

[0028] Optionally, S2 includes the following steps:

[0029] S21. Establish the overall resource pool model of the AI ​​server and define the total computing resources R of the AI ​​server. C , total storage resources R M and the total amount of training resources R T , and monitor and update each total amount in real time;

[0030] S22. Construct a macro-level recursive game network model based on different AI models carried by the AI ​​server, and define each AI model as a game agent A. j ;

[0031] S23. For each agent A j Define its resource demand vector Q j :

[0032]

[0033] in, Represents agent A j The computing resource requirements, Represents agent A j The storage resource requirements Represents agent A j The amount of training resources required;

[0034] S24. Define Agent A j The utility function U j (t), used to evaluate its satisfaction under the strategy combination:

[0035]

[0036] Among them, β j For Agent A j The task priority weight reflects the importance of the AI ​​model in server resource allocation, γ C , γ M , γ T is the weight coefficient of computing resources, data storage resources and training resources, satisfying γ C +γ M +γ T =1, For Agent A j The maximum resource requirement is set based on historical data and task requirements, λ j For Agent Aj The penalty coefficient, Ψ j (t) is the penalty function for resource over-application:

[0037]

[0038] Among them, η C , η M , η T Set the upper limit of reasonable usage of server resources for the threshold ratio of computing resources, data storage resources and training resources, (·) + Indicates taking a positive value, that is, when the value in the brackets is greater than zero, take the value, otherwise take zero;

[0039] S26, through the iterative calculation of the macro-level recursive game network model, agent A j Solve the optimization problem to maximize its utility function U by integrating the strategies of other agents j (t):

[0040]

[0041] Constraints:

[0042] Among them, the constraint condition is that the total resource application of all agents does not exceed the total available resources of the AI ​​server. The optimal resource allocation strategy is finally obtained:

[0043]

[0044] Optionally, S4 includes the following steps:

[0045] S41. Under the resource allocation strategy generated by the macro-level recursive game network model, the specific computing task module, data storage module and training task module in the AI ​​server are defined as micro-agent B respectively. k , each agent B k The resource requirement vector is defined as:

[0046]

[0047] in, Represents the micro-agent B k The computing resource requirements, Represents the micro-agent B k The data storage resource requirements, Represents the micro-agent B k The amount of training resources required;

[0048] S42, is the micro-level intelligent agent B k Allocate resource quota Pk , the resource quota is provided by the optimal allocation strategy at the macro level;

[0049] S43. Each micro-agent B k The utility function U k (t) To measure the degree of optimization of resource allocation:

[0050]

[0051] Among them, δ k For Agent B k The task priority weight, κ C , κ M , κ T is the utility weight coefficient of computing resources, storage resources and training resources, satisfying κ C +κ M +κ T =1,ξ k For Agent B k The penalty coefficient controls the negative effects of unreasonable applications in the resource allocation process, Φ k (t) is the resource over-application penalty function, resource over-application penalty function Φ k (t) is used to evaluate agent B k Whether the resource allocation of exceeds its allocation limit:

[0052]

[0053] in,(·) + Indicates a positive value. If agent B k If the demand exceeds the allocated resources, the value is greater than zero, indicating that a penalty is required;

[0054] S45, through the micro-level recursive game network model, agent B performs iterative calculations k Solve the optimization problem in combination with other agent strategies:

[0055]

[0056] Constraints:

[0057] Get Agent B k Optimized resource allocation strategy:

[0058]

[0059] S46. Optimal allocation strategy for calculating micro-level recursive game network models It is applied to the resource control of specific task modules in the AI ​​server, updates the resource usage of each module in real time, and feeds back to the macro-level recursive game network model for resource reallocation or optimization adjustment.

[0060] Optionally, the S6 includes the following steps:

[0061] S61, Micro-level recursive game network model for each micro-agent B k Monitor the current resource usage status in real time and compare its optimization resource requirements and resource quota P k (t), if it is detected that a subtask has insufficient resources or the load exceeds the current quota, the resource demand adjustment request process is triggered;

[0062] S62. Micro-level recursive game network model constructs resource adjustment requests based on resource demands that exceed quotas

[0063]

[0064] in: Represents agent B k The amount of computing resources, storage resources, and training resources that are currently required.

[0065] S63: Submit the resource adjustment request to the macro-level recursive game network model. The macro-level model aggregates the resource adjustment requests of all micro-agents to form a total demand adjustment set R adj (t):

[0066]

[0067] Among them, p is the total number of micro-agents;

[0068] S64, macro-level recursive game network model receives the total demand adjustment set R adj (t), the resource allocation strategy is re-optimized according to the current global resource surplus R(t) and load priority;

[0069] S65. According to the new resource allocation strategy generated by the macro-level model, the adjusted resource quota It is sent to the micro-level recursive game network model, while recording the current resource allocation status and updating the global resource pool in real time:

[0070]

[0071] S66. Micro-level recursive game network model after receiving the adjusted resource quota Then recalculate each agent B k The optimal resource allocation strategy

[0072]

[0073] The actual resource usage of each task module is updated in real time. If the resource shortage is triggered again in the future, the adjustment process of S61-S66 is repeated until the resource allocation is balanced.

[0074] Optionally, the S8 includes the following steps:

[0075] S81. Real-time monitoring of macro-level agent A during the operation of the AI ​​server j and micro-level agent B k The resource allocation and usage status of each time point t is and Record to historical data collection

[0076] S82. Calculate the error function E(t) between the actual resource usage and allocation of the macro- and micro-level game agents based on the real-time load data and historical data:

[0077]

[0078] in, A is a macro-level agent j The actual resource usage, For the micro-level agent B k The actual resource usage of , ∥·∥ represents the bi-norm of the vector, which is used to quantify the difference between allocation and usage;

[0079] S83. The utility functions of the macro-level recursive game network model and the micro-level recursive game network model are adjusted, and an error term is added to optimize the resource regulation strategy. The corrected macro-agent utility function is:

[0080] U′ j (t) = U j (t)-αE(t);

[0081] The modified micro-agent utility function is:

[0082] U′ k (t) = U k (t)-βE(t);

[0083] Among them, α and β are error penalty coefficients, which are used to balance the impact of resource usage errors on utility;

[0084] S84. Based on the revised utility function, a recursive feedback algorithm is used to adjust the agent's decision-making strategy and update the game strategy set for the next round of resource allocation. and

[0085] S85. Iteratively optimize the decision-making behaviors of the macro-level game agent and the micro-level game agent using the updated game strategy set, so that resource allocation gradually tends to the optimal state;

[0086] S86. Iterate and optimize the resource allocation strategy and It is continuously applied to macro-level recursive game network models and micro-level recursive game network models, and dynamically updates the real-time resource usage status to form a closed-loop optimization process.

[0087] The beneficial effects of the present invention are:

[0088] (1) The present invention adopts a hierarchical structure that combines macro-level and micro-level recursive game networks to achieve refined management of resource allocation from global to local. The macro-level recursive game network model is used to dynamically optimize the allocation of the overall resource pool of the AI ​​server, and the resource requirements of the task module are regulated in real time in combination with the micro-level recursive game network model, so that computing resources, storage resources and training resources can be dynamically adjusted according to different load fluctuation scenarios. Compared with the traditional single-layer balancing strategy, the present invention can effectively avoid resource waste and allocation imbalance caused by fixed rules or a single optimization target. Experiments show that the resource utilization efficiency of the present invention is improved by more than 15% in multi-task high-fluctuation load scenarios.

[0089] (2) The present invention introduces a recursive optimization mechanism based on error feedback to calculate the error function in task resource allocation and incorporates it into the utility functions at the macro and micro levels. The decision-making strategy of the intelligent agent is optimized in real time through a recursive feedback algorithm so that resource allocation gradually tends to the optimal in multiple iterations. Unlike traditional methods that only rely on real-time load data for adjustment, the present invention combines historical data to predict future load trends and adjust resource allocation strategies in advance, thereby effectively alleviating the system performance bottleneck problem caused by load surges. Test results show that in response to high load fluctuations, the resource response time of the present invention is shortened by more than 20%, and the system performance stability is significantly improved.

[0090] (3) The present invention balances the fairness of resource allocation and the priority of key tasks by introducing a resource over-application penalty function and a dynamic priority game mechanism. If the agent's resource application exceeds the current allocation quota, the penalty function will be triggered to suppress excessive resource application. At the same time, the resource allocation weight is dynamically adjusted according to the task priority to ensure that high-priority tasks are fully guaranteed when resources are tight. Compared with the problem that traditional load balancing methods are difficult to distinguish task priorities, the present invention achieves accurate satisfaction of high-priority task resource requirements through an intelligent priority control mechanism, and the key task completion rate is improved by 12%, effectively avoiding task interruption or delay caused by resource competition. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] 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:

[0092] Figure 1 This is a flow chart of a method for self-balancing load fluctuations of a storage and computing training AI server based on a recursive game network proposed by the present invention. DETAILED DESCRIPTION

[0093] 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.

[0094] refer to Figure 1 , a method for self-balancing load fluctuation of storage and computing training AI server based on recursive game network, comprising the following steps:

[0095] S1. Monitor each computing task in the AI ​​server, collect the computing resource requirements, data storage resource requirements, and training resource requirements of each computing task in real time, and record the real-time load data of each task;

[0096] S2. Construct a macro-level recursive game network model based on the overall resource pool of the AI ​​server, take the different AI models carried by the AI ​​server as game agents, and allocate and control the total amount of computing resources, storage resources, and training resources of the AI ​​server;

[0097] S3. Generate an overall resource allocation strategy through multi-agent competitive calculation of a macro-level recursive game network model, determine the initial computing resource allocation, storage resource allocation, and training resource allocation for each AI model based on the task priority and resource requirements of each AI model in the server, and dynamically respond to the prediction of load fluctuations;

[0098] S4. Based on the resource allocation strategy generated at the macro level, a micro-level recursive game network model is constructed. The specific computing task module, data storage module, and training task module in the AI ​​server are defined as micro-agents. Under the resource quota provided by the macro level, each agent plays a game based on the resource requirements and load fluctuations of the task, and achieves real-time load regulation and resource allocation for each subtask through competitive behavior.

[0099] S5. In the micro-level recursive game network model, each micro-agent analyzes the load of the current subtask in real time, calculates the resource occupancy rate and idle rate, and inputs the feedback results into the micro-level recursive game network model. Each micro-agent dynamically adjusts the resource usage ratio according to the feedback resource usage and task load changes through an adaptive control mechanism;

[0100] S6. Combining the game results of the macro-level recursive game network model and the micro-level recursive game network model, when it is detected that a subtask has insufficient resources or the load exceeds the current quota, the micro-level recursive game network model submits a resource demand adjustment request to the macro-level recursive game network model, and the macro-level recursive game network model reallocates the overall resources of the AI ​​model and sends the adjusted resource quota to the micro-level recursive game network model in real time;

[0101] S7. Based on the historical load data and the real-time load data in the macro-level recursive game network model and the micro-level recursive game network model, the future load fluctuation trend is predicted through a recursive feedback model to generate a trend estimate of resource demand;

[0102] S8. During the load balancing operation of the AI ​​server, dynamically adjust the resource control strategies of the macro-level recursive game network model and the micro-level recursive game network model, gradually optimize the resource decisions of the game agents, and make the resource allocation gradually tend to the optimal state.

[0103] In this implementation, S1 includes the following steps:

[0104] S11. Monitor each computing task in the AI ​​server in real time, collect computing resource requirements, data storage resource requirements, and training resource requirements of the computing task, and the real-time load data of the computing task is expressed as L i (t), where i is the unique identifier of the task and t is the time point;

[0105] S12. Record the real-time computing resource requirements, data storage resource requirements, and training resource requirements of each computing task respectively, and obtain the comprehensive load data vector D of each task. i (t):

[0106] D i (t) = {Ci (t),M i (t),T i (t)};

[0107] Among them, C i (t) represents the amount of computing resources required by task i at time t, in floating-point operations, M i (t) represents the amount of data storage resources required by task i at time t, in bytes, T i (t) represents the amount of training resources required for task i at time t, in training steps;

[0108] S13, aggregating the real-time load data of multiple tasks at time point t to form a time series task load data set D(t);

[0109] S14. Record the load change trend of each task and calculate the load change rate Δ of task i i (t):

[0110]

[0111] Wherein, Δt is the time interval of load collection;

[0112] S15, input the real-time collected task load data, comprehensive load data vector and load change rate into the dynamic load data storage module to form a task real-time load data set

[0113]

[0114] Among them, Δ(t)={Δ1(t),Δ2(t),…,Δ n (t)} is the set of collected time points.

[0115] In this implementation, S2 includes the following steps:

[0116] S21. Establish the overall resource pool model of the AI ​​server and define the total computing resources R of the AI ​​server. C , total storage resources R M and the total amount of training resources R T , and monitor and update each total amount in real time;

[0117] S22. Construct a macro-level recursive game network model based on different AI models carried by the AI ​​server, and define each AI model as a game agent A. j ;

[0118] S23. For each agent A j Define its resource demand vector Q j :

[0119]

[0120] in, Represents agent A j The computing resource requirements, Represents agent A j The storage resource requirements Represents agent A j The amount of training resources required;

[0121] S24. Define Agent A j The utility function U j (t), used to evaluate its satisfaction under the strategy combination:

[0122]

[0123] Among them, β j For Agent A j The task priority weight reflects the importance of the AI ​​model in server resource allocation, γ C , γ M , γ T is the weight coefficient of computing resources, data storage resources and training resources, satisfying γ C +γ M +γ T =1, For Agent A j The maximum resource requirement is set based on historical data and task requirements, λ j For Agent A j The penalty coefficient, Ψ j (t) is the penalty function for resource over-application:

[0124]

[0125] Among them, η C , η M , η T Set the upper limit of reasonable usage of server resources for the threshold ratio of computing resources, data storage resources and training resources, (·) + Indicates taking a positive value, that is, when the value in the brackets is greater than zero, take the value, otherwise take zero;

[0126] S26, through the iterative calculation of the macro-level recursive game network model, agent A j Solve the optimization problem to maximize its utility function U by integrating the strategies of other agents j (t):

[0127]

[0128] Constraints:

[0129] Among them, the constraint condition is that the total resource application of all agents does not exceed the total available resources of the AI ​​server. The optimal resource allocation strategy is finally obtained:

[0130]

[0131] In this implementation, S4 includes the following steps:

[0132] S41. Under the resource allocation strategy generated by the macro-level recursive game network model, the specific computing task module, data storage module and training task module in the AI ​​server are defined as micro-agent B respectively. k , each agent B k The resource demand vector is defined as:

[0133]

[0134] in, Represents the micro-agent B k The computing resource requirements, Represents the micro-agent B k The data storage resource requirements, Represents the micro-agent B k The amount of training resources required;

[0135] S42, is the micro-level intelligent agent B k Allocate resource quota P k , the resource quota is provided by the optimal allocation strategy at the macro level;

[0136] S43. Each micro-agent B k The utility function U k (t) To measure the degree of optimization of resource allocation:

[0137]

[0138] Among them, δ k For Agent B k The task priority weight, κ C , κ M , κ T is the utility weight coefficient of computing resources, storage resources and training resources, satisfying κ C +κ M +κ T =1,ξ k For Agent B k The penalty coefficient controls the negative effects of unreasonable applications in the resource allocation process, Φ k(t) is the resource over-application penalty function, resource over-application penalty function Φ k (t) is used to evaluate agent B k Whether the resource allocation of exceeds its allocation limit:

[0139]

[0140] in,(·) + Indicates a positive value. If agent B k If the demand exceeds the allocated resources, the value is greater than zero, indicating that a penalty is required;

[0141] S45, through the micro-level recursive game network model, agent B performs iterative calculations k Solve the optimization problem in combination with other agent strategies:

[0142]

[0143] Constraints:

[0144] Get Agent B k Optimized resource allocation strategy:

[0145]

[0146] S46. Optimal allocation strategy for calculating micro-level recursive game network models It is applied to the resource control of specific task modules in the AI ​​server, updates the resource usage of each module in real time, and feeds back to the macro-level recursive game network model for resource reallocation or optimization adjustment.

[0147] In this implementation, S6 includes the following steps:

[0148] S61, Micro-level recursive game network model for each micro-agent B k Monitor the current resource usage status in real time and compare its optimization resource requirements and resource quota P k (t), if it is detected that a subtask has insufficient resources or the load exceeds the current quota, the resource demand adjustment request process is triggered;

[0149] S62. Micro-level recursive game network model constructs resource adjustment requests based on resource demands that exceed quotas

[0150]

[0151] in: Represents agent B kThe amount of computing resources, storage resources, and training resources that are currently required.

[0152] S63: Submit the resource adjustment request to the macro-level recursive game network model. The macro-level model aggregates the resource adjustment requests of all micro-agents to form a total demand adjustment set R adj (t):

[0153]

[0154] Among them, p is the total number of micro-agents;

[0155] S64, macro-level recursive game network model receives the total demand adjustment set R adj (t), the resource allocation strategy is re-optimized according to the current global resource surplus R(t) and load priority;

[0156] S65. According to the new resource allocation strategy generated by the macro-level model, the adjusted resource quota It is sent to the micro-level recursive game network model, while recording the current resource allocation status and updating the global resource pool in real time:

[0157]

[0158] S66. Micro-level recursive game network model after receiving the adjusted resource quota Then recalculate each agent B k The optimal resource allocation strategy

[0159]

[0160] The actual resource usage of each task module is updated in real time. If the resource shortage is triggered again in the future, the adjustment process of S61-S66 is repeated until the resource allocation is balanced.

[0161] In this implementation, S8 includes the following steps:

[0162] S81. Real-time monitoring of macro-level agent A during the operation of the AI ​​server j and micro-level agent B k The resource allocation and usage status of each time point t is and Record to historical data collection

[0163] S82. Calculate the error function E(t) between the actual resource usage and allocation of the macro- and micro-level game agents based on the real-time load data and historical data:

[0164]

[0165] in, A is a macro-level agent j The actual resource usage, For the micro-level agent B k The actual resource usage of , ∥·∥ represents the bi-norm of the vector, which is used to quantify the difference between allocation and usage;

[0166] S83. The utility functions of the macro-level recursive game network model and the micro-level recursive game network model are adjusted, and an error term is added to optimize the resource regulation strategy. The corrected macro-agent utility function is:

[0167] U′ j (t) = U j (t)-αE(t);

[0168] The modified micro-agent utility function is:

[0169] U′ k (t) = U k (t)-βE(t);

[0170] Among them, α and β are error penalty coefficients, which are used to balance the impact of resource usage errors on utility;

[0171] S84. Based on the revised utility function, a recursive feedback algorithm is used to adjust the agent's decision-making strategy and update the game strategy set for the next round of resource allocation. and

[0172] S85. Iteratively optimize the decision-making behaviors of the macro-level game agent and the micro-level game agent using the updated game strategy set, so that resource allocation gradually tends to the optimal state;

[0173] S86. Iterate and optimize the resource allocation strategy and It is continuously applied to macro-level recursive game network models and micro-level recursive game network models, and dynamically updates the real-time resource usage status to form a closed-loop optimization process.

[0174] Embodiment 1:

[0175] The present invention adopts a hierarchical structure that combines macro-level and micro-level recursive game networks to achieve refined management of resource allocation from global to local. The macro-level recursive game network model is used to dynamically optimize the allocation of the overall resource pool of the AI ​​server, and the resource requirements of the task module are regulated in real time in combination with the micro-level recursive game network model, so that computing resources, storage resources and training resources can be dynamically adjusted for different load fluctuation scenarios. Compared with traditional single-layer balancing strategies, the present invention can effectively avoid resource waste and allocation imbalance caused by fixed rules or a single optimization target. Experiments show that the resource utilization efficiency of the present invention is improved by more than 15% in multi-task and high-fluctuation load scenarios.

[0176] The present invention introduces a recursive optimization mechanism based on error feedback to calculate the error function in task resource allocation and incorporates it into the utility functions at the macro and micro levels. The decision-making strategy of the intelligent agent is optimized in real time through the recursive feedback algorithm so that resource allocation gradually tends to the optimal in multiple iterations. Unlike the traditional method that only relies on real-time load data for adjustment, the present invention combines historical data to predict future load trends and adjust the resource allocation strategy in advance, thereby effectively alleviating the system performance bottleneck problem caused by load surges. Test results show that in response to high load fluctuations, the resource response time of the present invention is shortened by more than 20%, and the system performance stability is significantly improved.

[0177] The present invention balances the fairness of resource allocation and the priority of key tasks by introducing a resource over-application penalty function and a dynamic priority game mechanism. If the agent's resource application exceeds the current allocation quota, the penalty function will be triggered to suppress excessive resource application. At the same time, the resource allocation weight is dynamically adjusted according to the task priority to ensure that high-priority tasks are fully guaranteed when resources are tight. Compared with the problem that traditional load balancing methods are difficult to distinguish task priorities, the present invention achieves accurate satisfaction of high-priority task resource requirements through an intelligent priority control mechanism, and the key task completion rate is increased by 12%, effectively avoiding task interruptions or delays caused by resource competition.

[0178] 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 self-balancing load fluctuations of storage and computing training AI servers based on recursive game networks, characterized in that: The steps include: S1. Monitor each computing task in the AI ​​server, collect the computing resource requirements, data storage resource requirements, and training resource requirements of each computing task in real time, and record the real-time load data of each task; S2. Construct a macro-level recursive game network model based on the overall resource pool of the AI ​​server, take the different AI models carried by the AI ​​server as game agents, and allocate and control the total amount of computing resources, storage resources, and training resources of the AI ​​server; S3. Generate an overall resource allocation strategy through multi-agent competitive calculation of a macro-level recursive game network model, determine the initial computing resource allocation, storage resource allocation, and training resource allocation for each AI model based on the task priority and resource requirements of each AI model in the server, and dynamically respond to the prediction of load fluctuations; S4. Based on the resource allocation strategy generated at the macro level, a micro-level recursive game network model is constructed. The specific computing task module, data storage module, and training task module in the AI ​​server are defined as micro-agents. Under the resource quota provided by the macro level, each agent plays a game based on the resource requirements and load fluctuations of the task, and achieves real-time load regulation and resource allocation for each subtask through competitive behavior. S5. In the micro-level recursive game network model, each micro-agent analyzes the load of the current subtask in real time, calculates the resource occupancy rate and idle rate, and inputs the feedback results into the micro-level recursive game network model. Each micro-agent dynamically adjusts the resource usage ratio according to the feedback resource usage and task load changes through an adaptive control mechanism; S6. Combining the game results of the macro-level recursive game network model and the micro-level recursive game network model, when it is detected that a subtask has insufficient resources or the load exceeds the current quota, the micro-level recursive game network model submits a resource demand adjustment request to the macro-level recursive game network model, and the macro-level recursive game network model reallocates the overall resources of the AI ​​model and sends the adjusted resource quota to the micro-level recursive game network model in real time; S7. Based on the historical load data and the real-time load data in the macro-level recursive game network model and the micro-level recursive game network model, the future load fluctuation trend is predicted through a recursive feedback model to generate a trend estimate of resource demand; S8. During the load balancing operation of the AI ​​server, dynamically adjust the resource control strategies of the macro-level recursive game network model and the micro-level recursive game network model, gradually optimize the resource decisions of the game agents, and make the resource allocation gradually tend to the optimal state.

2. According to claim 1, a method for self-balancing the load fluctuation of storage and computing training AI servers based on a recursive game network is characterized in that: The S1 comprises the following steps: S11. Monitor each computing task in the AI ​​server in real time, collect the computing resource requirements, data storage resource requirements, and training resource requirements of the computing task, and the real-time load data of the computing task is expressed as Li(t), where i is the unique identifier of the task and t is the time point; S12. Record the real-time computing resource requirements, data storage resource requirements, and training resource requirements of each computing task respectively, and obtain the comprehensive load data vector D of each task. i (t): D i (t)={C i (t),M i (t),T i (t)}; Among them, C i (t) represents the amount of computing resources required by task i at time t, in floating-point operations, M i (t) represents the amount of data storage resources required by task i at time t, in bytes, and Ti(t) represents the amount of training resources required by task i at time t, in training steps; S13, aggregating the real-time load data of multiple tasks at time point t to form a time series task load data set D(t); S14. Record the load change trend of each task and calculate the load change rate Δ of task i i (t): Wherein, Δt is the time interval of load collection; S15, input the real-time collected task load data, comprehensive load data vector and load change rate into the dynamic load data storage module to form a task real-time load data set D: Among them, Δ(t)={Δ1(t), Δ2(t),..., Δ n (t)} is the set of collected time points.

3. According to claim 1, a method for self-balancing the load fluctuation of storage and computing training AI servers based on a recursive game network is characterized in that: The S2 comprises the following steps: S21. Establish the overall resource pool model of the AI ​​server and define the total computing resources RC and storage resources R of the AI ​​server. M and total amount of training resources RT, and monitor and update each total amount in real time; S22. Construct a macro-level recursive game network model based on different AI models carried by the AI ​​server, and define each AI model as a game agent A. j ; S23. For each agent A j Define its resource demand vector Q j : in, Represents agent A j The computing resource requirements, Represents agent A j The storage resource requirements Represents agent A i The amount of training resources required; S24. Define Agent A j The utility function U j (t), used to evaluate its satisfaction under the strategy combination: Among them, β j For Agent A j The task priority weight reflects the importance of the AI ​​model in server resource allocation, γ C , γ M , γ T is the weight coefficient of computing resources, data storage resources and training resources, satisfying γ C +γ M +γ T =1, For Agent A j The maximum resource requirement is set based on historical data and task requirements, λ j For Agent A j The penalty coefficient, Ψ j (t) is the penalty function for resource over-application: Among them, η C , η M , η T Set the reasonable upper limit of server resource usage for the threshold ratio of computing resources, data storage resources, and training resources. (·) + Indicates taking a positive value, that is, when the value in the brackets is greater than zero, take the value, otherwise take zero; S26, through the iterative calculation of the macro-level recursive game network model, agent A j Solve the optimization problem to maximize its utility function U by integrating the strategies of other agents j (t): Constraints: Among them, the constraint condition is that the total resource application of all agents does not exceed the total available resources of the AI ​​server. The optimal resource allocation strategy is finally obtained:

4. According to claim 1, a method for self-balancing the load fluctuation of storage and computing training AI servers based on a recursive game network, characterized in that: The S4 comprises the following steps: S41. Under the resource allocation strategy generated by the macro-level recursive game network model, the specific computing task module, data storage module and training task module in the AI ​​server are defined as micro-agent B respectively. k , each agent B k The resource requirement vector is defined as: in, Represents the micro-agent B k The computing resource requirements, Represents the micro-agent B k The data storage resource requirements, Represents the micro-agent B k The amount of training resources required; S42, is the micro-level intelligent agent B k Allocate resource quota P k , the resource quota is provided by the optimal allocation strategy at the macro level; S43. Each micro-agent B k The utility function U k (t) To measure the degree of optimization of resource allocation: Among them, δ k For agent B k The task priority weight, κ C , κ M , κ T is the utility weight coefficient of computing resources, storage resources and training resources, satisfying κ C +κ M +κ T =1,ξ k For agent B k The penalty coefficient controls the negative effects of unreasonable applications in the resource allocation process, Φp k (t) is the resource over-application penalty function, resource over-application penalty function Φ k (t) is used to evaluate agent B k Whether the resource allocation of exceeds its allocation limit: in,(·) + Indicates a positive value. If agent B k If the demand exceeds the allocated resources, the value is greater than zero, indicating that a penalty is required; S45, through the micro-level recursive game network model, agent B performs iterative calculations k Solve the optimization problem in combination with other agent strategies: Constraints: Get Agent B k Optimized resource allocation strategy: S46. Optimal allocation strategy for calculating micro-level recursive game network models It is applied to the resource control of specific task modules in the AI ​​server, updates the resource usage of each module in real time, and feeds back to the macro-level recursive game network model for resource reallocation or optimization adjustment.

5. According to claim 1, a method for self-balancing load fluctuations of storage and computing training AI servers based on recursive game networks is characterized in that: The S6 comprises the following steps: S61, Micro-level recursive game network model for each micro-agent B k Monitor the current resource usage status in real time and compare its optimization resource requirements and resource quota P k (t), if it is detected that a subtask has insufficient resources or the load exceeds the current quota, the resource demand adjustment request process is triggered; S62. Micro-level recursive game network model constructs resource adjustment requests based on resource demands that exceed quotas in: Represents agent B k The amount of computing resources, storage resources, and training resources that are currently required. S63: Submit the resource adjustment request to the macro-level recursive game network model. The macro-level model aggregates the resource adjustment requests of all micro-agents to form a total demand adjustment set R adj (t): Among them, p is the total number of micro-agents; S64, macro-level recursive game network model receives the total demand adjustment set R adj (t), the resource allocation strategy is re-optimized according to the current global resource surplus R(t) and load priority; S65. According to the new resource allocation strategy generated by the macro-level model, the adjusted resource quota It is sent to the micro-level recursive game network model, while recording the current resource allocation status and updating the global resource pool in real time: S66. Micro-level recursive game network model after receiving the adjusted resource quota Then recalculate each agent B k The optimal resource allocation strategy The actual resource usage of each task module is updated in real time. If the resource shortage is triggered again in the future, the adjustment process of S61-S66 is repeated until the resource allocation is balanced.

6. According to claim 1, a method for self-balancing load fluctuations of storage and computing training AI servers based on recursive game networks is characterized in that: The S8 comprises the following steps: S81. Real-time monitoring of macro-level agent A during the operation of the AI ​​server j and micro-level agent B k The resource allocation and usage status of each time point t is and Record to historical data collection S82. Calculate the error function E(t) between the actual resource usage and allocation of the macro- and micro-level game agents based on the real-time load data and historical data: in, A is a macro-level agent j The actual resource usage, For the micro-level agent B k The actual resource usage of , ∥·∥ represents the bi-norm of the vector, which is used to quantify the difference between allocation and usage; S83. The utility functions of the macro-level recursive game network model and the micro-level recursive game network model are adjusted, and an error term is added to optimize the resource regulation strategy. The corrected macro-agent utility function is: U′ j (t)=U j (t)-αE(t); The modified micro-agent utility function is: U′ k (t)=U k (t)-βE(t); Among them, α and β are error penalty coefficients, which are used to balance the impact of resource usage errors on utility; S84. Based on the revised utility function, a recursive feedback algorithm is used to adjust the agent's decision-making strategy and update the game strategy set for the next round of resource allocation. and S85. Iteratively optimize the decision-making behaviors of the macro-level game agent and the micro-level game agent using the updated game strategy set, so that resource allocation gradually tends to the optimal state; S86. Iterate and optimize the resource allocation strategy and It is continuously applied to macro-level recursive game network models and micro-level recursive game network models, and dynamically updates the real-time resource usage status to form a closed-loop optimization process.

Images