Game theory-based improved model cloud computing node storage, calculation and training resource competition scheduling method

Abstract

The invention discloses a game theory improved model-based cloud computing node storage calculation training resource competition scheduling method. The method comprises the following steps of S1, constructing a global resource scheduling system; s2, carrying out initial division on the cloud computing nodes in each logic partition; s3, executing a local resource competition game model; s4, the global resource scheduling system collects a local resource competition result of each logic partition and residual available resources in each logic partition; s5, executing the global collaborative game model; s6, dynamically updating the resource state information of the cloud computing node; s7, the global resource scheduling system performs adaptive adjustment on resource demand strategies in the local game model and the global cooperative game model; and S8, the global resource scheduling system stores and analyzes past resource competition results and task execution conditions, and each cloud computing node updates a resource demand strategy according to the data. According to the invention, the problems of resource bottleneck and overload in the prior art are solved, and the resource utilization efficiency and the scheduling flexibility are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing technology, and in particular to a method for scheduling storage, computing and training resource competition of cloud computing nodes based on an improved model of game theory. Background Art

[0002] In the prior art, resource scheduling of cloud computing nodes mainly relies on pre-set static rules or simple load balancing strategies. Existing strategies are prone to uneven resource allocation when facing complex cloud computing environments with multiple nodes and multiple tasks. Task execution efficiency is significantly reduced when there is fierce competition for storage, computing and training resources. In traditional methods, cloud computing systems often perform resource scheduling through static allocation or first-come, first-served methods, which fails to fully consider the real-time resource load and task demand differences of each node, resulting in frequent resource bottlenecks and overloads.

[0003] For example, current scheduling strategies mainly focus on optimizing a single resource, but often ignore the synchronous scheduling requirements of storage resources and training resources. In cloud computing environments that require a large number of parallel computing and deep learning tasks, the demand for resources for training tasks fluctuates greatly. Traditional scheduling systems are difficult to dynamically adjust resource allocation according to real-time task loads, which can easily lead to excessive consumption of computing resources and insufficient training resources. This resource scheduling method not only reduces resource utilization efficiency, but may also cause tasks to fail to be completed on time, thereby affecting the overall system performance.

[0004] In addition, most resource scheduling algorithms in the existing technology lack intelligence and adaptability. When the task load changes greatly, the traditional scheduling system cannot respond in real time, resulting in excessive resource utilization of some nodes, while the idle resources of other nodes further aggravate the waste of resources and task delays. In the multi-node and multi-resource scheduling scenario, the lack of an effective coordination mechanism between nodes makes it impossible to achieve global optimization of resources, resulting in poor performance of the system when processing high-concurrency tasks.

[0005] In summary, the existing technologies have the following main shortcomings: First, resource allocation is uneven. Especially in scenarios where storage, computing and training resources compete at the same time, traditional methods cannot effectively solve resource bottlenecks and overload problems. Second, existing scheduling methods are insufficient in the ability to optimize the synchronization of multiple resources, which can easily lead to uneven resource utilization and affect the overall performance of the system. Finally, there is a lack of intelligent and adaptive scheduling mechanisms, and the scheduling strategy cannot be dynamically adjusted according to task requirements and resource status, resulting in poor performance of the system when processing high-concurrency tasks. Summary of the invention

[0006] One purpose of the present invention is to propose a method for scheduling storage, computing and training resource competition of cloud computing nodes based on an improved model of game theory. The present invention solves the problems of resource bottleneck and overload in the prior art and greatly improves resource utilization efficiency and scheduling flexibility.

[0007] According to an embodiment of the present invention, a method for scheduling storage, computing and training resource competition of cloud computing nodes based on an improved model of game theory includes the following steps:

[0008] S1. In a cloud computing environment, deploy multiple cloud computing nodes distributed in multiple logical partitions, each cloud computing node includes storage resources, computing resources, and training resources, and build a global resource scheduling system;

[0009] S2. Initially divide the cloud computing nodes in each logical partition, collect resource status information of each cloud computing node, and build a local resource competition game model in the logical partition based on the collected resource status information. Each cloud computing node in the logical partition is regarded as an independent game participant, and the nodes formulate resource competition strategies according to their own task requirements, resource load and task priority.

[0010] S3. Execute the local resource competition game model. Through the strategy game process, the cloud computing nodes in the logical partition adjust the resource demand strategy, optimize the use of storage resources, computing resources and training resources, generate local resource competition results, and allocate resources in the logical partition;

[0011] S4. The global resource scheduling system collects the local resource competition results of each logical partition and the remaining available resources in each logical partition. Based on the global collaborative game model in game theory, each logical partition is regarded as a global game participant. Each logical partition adjusts its resource usage strategy according to the local resource competition results and global resource demand.

[0012] S5. Execute the global collaborative game model. Under the global coordination of the global resource scheduling system, each logical partition performs optimal scheduling of global resources through strategy adjustment and information sharing, generates a global resource allocation plan, and coordinates the allocation of storage resources, computing resources, and training resources among the logical partitions.

[0013] S6. The global resource scheduling system collects the resource usage and task execution status of each cloud computing node in real time through the monitoring module, and dynamically updates the resource status information of the cloud computing node according to the dynamic changes of task load, resource bottleneck and global performance of the system;

[0014] S7, the global resource scheduling system adaptively adjusts the resource demand strategies in the local game model and the global collaborative game model, triggers the global resource scheduling system to perform global scheduling optimization when a resource bottleneck or task overload in a logical partition is detected, and prioritizes the allocation of storage resources, computing resources, and training resources from other logical partitions;

[0015] S8. Strategy optimization and learning are performed based on historical game data. The global resource scheduling system stores and analyzes past resource competition results and task execution status. Each cloud computing node updates its resource demand strategy based on the data and makes better resource allocation decisions in future game processes.

[0016] Optionally, the S1 further comprises:

[0017] S11. Construct a physical cluster including multiple logical partitions in a cloud computing environment. Each logical partition is composed of a number of cloud computing nodes including storage resources, computing resources, and training resources. The distribution of each cloud computing node in different logical partitions depends on the network topology and task load requirements of each logical partition.

[0018] S12: Initialize and configure the storage resources, computing resources, and training resources in the cloud computing nodes. i Represents each cloud computing node N i The initial available storage capacity and computing resources C i Represents each cloud computing node N i The initial available computing power, training resources T i Represents each cloud computing node N i Initial available training capacity;

[0019] S13, according to each cloud computing node N i The geographical location, network bandwidth and task requirements determine the node distribution within each logical partition, so that the storage resources, computing resources and training resources within the logical partition are evenly distributed, forming a global resource scheduling system structure:

[0020] S i +C i +T i =R i ;

[0021] Among them, R i Represents the cloud computing node N i The total resources, S i , C i and T i They represent the initial configuration amounts of storage resources, computing resources, and training resources respectively.

[0022] Optionally, the S2 further comprises:

[0023] S21, initially divide the cloud computing nodes in each logical partition, collect the N ij Resource status information, where N ij represents the i-th cloud computing node in logical partition j;

[0024] S22, collect cloud computing nodes N ij Resource status information, including:

[0025] Storage resource load:

[0026]

[0027] Computing resource load:

[0028]

[0029] Training resource load:

[0030]

[0031] in, Represents the cloud computing nodes N ij of used storage resources, computing resources, and training resources, Represents the cloud computing nodes N ij The total storage resource capacity, total computing resource capacity and total training resource capacity;

[0032] S23. Calculate the cloud computing node N ij Remaining capacity:

[0033] Remaining storage capacity:

[0034]

[0035] Calculate the remaining capacity of resources:

[0036]

[0037] Remaining capacity of training resources:

[0038]

[0039] S24. Obtain cloud computing node N ij The task execution progress P ij , indicating the completion percentage of the current task;

[0040] S25, based on the collected resource status information, construct a local resource competition game model within the logical partition j, and divide each cloud computing node N within the logical partition ijTreated as independent game players;

[0041] S26, cloud computing node N ij According to your own task requirements ij , Resource Load and the task priority U ij Formulate resource competition strategies ij , task requirement D ij Including the required storage resources Computing resources and training resources

[0042] S27. Define cloud computing node N ij The profit function of the local resource competition game model Π ij , used to evaluate the effect of resource competition strategy, the benefit function comprehensively considers resource utilization efficiency, task completion speed and task priority:

[0043]

[0044] Among them, α1, β1, and γ1 are weight coefficients, which reflect the influence of resource demand, task progress, and task priority in the benefit function.

[0045] Optionally, the S3 further includes:

[0046] S31, execute the local resource competition game model to allocate each cloud computing node N in the logical partition j ij As a game participant, each node in the local resource competition game model is based on its current resource demand strategy S ij Participate in resource competition, the goal of the game process is to optimize the resource usage of cloud computing nodes;

[0047] S32, each cloud computing node N ij In the game process, according to its own resource status information and profit function Π ij , through the strategy adjustment process, iteratively optimizes its resource demand strategy S ij , the direction of strategy adjustment is guided by the goal of maximizing the profit function:

[0048]

[0049] Nodes are based on resource load and task requirements D ij Its resource demand strategy S ij Make dynamic adjustments;

[0050] S33, execute multiple rounds of strategy game process, each cloud computing node updates its resource demand strategy S after each round of game. ij, so that the resource utilization efficiency of the node is gradually optimized:

[0051]

[0052] Among them, η1 is the learning rate, represents the resource demand strategy after the tth round of game;

[0053] S34, the node continues to adjust its strategy in the local resource competition game until the preset convergence condition is met. The specific condition is the profit function Π ij The increment is lower than the set threshold ∈:

[0054]

[0055] S35. After the convergence conditions are met, the node generates the final local resource competition result, which includes the final resource demand strategy S of each cloud computing node. ij and the actual allocated storage resources Computing resources and training resources

[0056] S36. Based on the local resource competition results, the cloud computing nodes are allocated resources in the logical partition j. The allocation strategy is based on the final resource demand strategy S of the node. ij As well as the remaining available resources in the logical partition, the resource requirements of each node can be optimally met.

[0057] Optionally, the S4 further includes:

[0058] S41, the global resource scheduling system collects the local resource competition results of each logical partition j, including each cloud computing node N ij Resource demand strategy S ij and actual allocated storage resources Computing resources and training resources

[0059] S42: The global resource scheduling system collects the remaining available resources in each logical partition, and calculates the remaining available storage resources of each logical partition j based on the dynamic changes of resources and the dependencies between nodes. Computing resources and training resources

[0060]

[0061] in, They represent the remaining storage resources, computing resources, and training resources of the logical partition j at time t, respectively. Represents the cloud computing nodes Nij The total storage resources, total computing resources and total training resource capacity at time t, Represents the cloud computing nodes N ij The storage resources, computing resources, and training resources actually allocated at time t, The cloud computing nodes N ij Dynamic weighting coefficients on storage resources, computing resources, and training resources are used to measure the rate of change of actual node usage resources, ΔT ij (t) represents the cloud computing node N ij The task execution time difference at time t, ΔT max represents the maximum task execution time difference of all nodes in logical partition j, which is used to standardize the time difference, F ij (t) is the dynamic adjustment factor, which represents the cloud computing node N ij Resource utilization adjustment coefficient at time t;

[0062] S43, based on the collected local resource competition results and the remaining available resources of each logical partition, the global resource scheduling system regards each logical partition j as a global game participant to build a global collaborative game model, in which each logical partition j adjusts the global resource demand strategy S according to its local resource competition results and remaining resource conditions j :

[0063]

[0064] Among them, α j and β j The adjustment coefficient for each logical partition reflects the balance between the remaining resources and the demand of each logical partition. The storage, computing and training resource requirements of each node respectively.

[0065] Optionally, the S5 further includes:

[0066] S51, executing the global collaborative game model, with each logical partition j as a global game participant, and the global resource scheduling system summarizes and analyzes the local resource competition results and remaining resource conditions of each logical partition, builds a global resource scheduling strategy, and performs resource sharing and coordination in a global scope;

[0067] S52, the global resource scheduling system uses the global resource demand strategy S of each logical partition j and remaining resources The resource allocation strategy of each logical partition is adjusted through multiple rounds of games. Each logical partition shares its current resource demand and remaining resource status during the game. The game goal is to achieve optimal scheduling of global resources through strategy adjustment.

[0068] S53. Profit function Π of the global cooperative game model j Defined as the balance between resource utilization efficiency and resource supply demand for each logical partition:

[0069]

[0070] Among them, α j1 and β j1 is the balance coefficient, which reflects the degree of attention paid by each logical partition to the remaining resources and task priorities. j represents the overall progress of task completion in logical partition j, They are the storage, computing, and training resource requirements of the node respectively;

[0071] S54, execute multiple rounds of global game, each logical partition adjusts its resource demand strategy based on the feedback information obtained during the global game process j , adjust the direction by maximizing the profit function Π of the global cooperative game model j OK, the specific adjustment strategy is:

[0072]

[0073] Among them, η2 is the learning rate, represents the resource demand strategy after the tth round of global game;

[0074] S55. When the global game model meets the convergence condition: the profit function Π j The increment of is lower than the set threshold ∈, and a global resource allocation plan is generated. The global resource allocation plan determines the final allocation of storage resources, computing resources, and training resources among each logical partition. The allocation result is optimized based on the resource demand and remaining resource situation of each logical partition during the game process.

[0075] Optionally, the S7 further includes:

[0076] S71. The global resource scheduling system monitors the resource usage of each logical partition in real time, collects the load and remaining capacity of the storage resources, computing resources, and training resources of each cloud computing node, and identifies resource bottlenecks or task overloads.

[0077] S72. When the task load in a certain logical partition exceeds the resource threshold, the adaptive adjustment mechanism of the local game model is triggered. Each cloud computing node adjusts its resource competition strategy according to the current resource load, task demand and task priority to cope with the shortage or bottleneck of local resources, and optimizes the resource demand through the iterative process of the local game model.

[0078] S73, the global resource scheduling system performs adaptive adjustment of the global collaborative game model for each logical partition. When a task overload or resource bottleneck is detected, the global game mechanism coordinates the resource flow between multiple logical partitions, and each partition readjusts its global resource demand strategy according to the current local game result;

[0079] S74. When a partition faces resource shortage, the global resource scheduling system prioritizes allocating available storage resources, computing resources, and training resources from other partitions. The partitions coordinate resource flows through global game to maximize global resource utilization efficiency.

[0080] S75, the local game model and the global collaborative game model realize adaptive adjustment in the process, and the global resource scheduling system continuously updates the resource status of each partition and performs dynamic optimization according to the real-time monitoring data, so that the resource usage of the global resource scheduling system is in the optimal state under different task load conditions;

[0081] S76. The global resource scheduling system further optimizes the resource demand strategies of nodes and partitions based on historical game data, and improves the local game model and the global collaborative game model through a continuous adaptive learning mechanism.

[0082] The beneficial effects of the present invention are:

[0083] (1) The present invention divides the resource competition of cloud computing nodes into two layers: local game and global collaborative game model. It can dynamically optimize storage, computing, and training resources within a partition and globally. The local game model is used to schedule resource competition among nodes within a partition to ensure optimal resource allocation within the partition. The global collaborative game model is based on the global resource status and the game results of each partition. It realizes cross-partition resource scheduling through multiple rounds of strategy adjustment. The two-layer game structure breaks through the limitations of traditional static resource scheduling, greatly improves resource allocation efficiency, avoids resource bottlenecks within a single partition, and reduces the idleness of resources in other partitions to ensure efficient utilization of global resources under load changes.

[0084] (2) The present invention adopts an adaptive adjustment strategy by real-time monitoring of the resource status in the cloud computing environment, so that each node can flexibly adjust the resource competition strategy when the task load changes. When a resource bottleneck or task overload occurs, the global scheduling optimization is automatically triggered through the adaptive adjustment model, and available storage, computing and training resources are preferentially allocated from other partitions. This solves the scheduling lag problem in the prior art and enables the system to quickly adjust the resource allocation plan when facing high-concurrency tasks and dynamic loads, thereby ensuring the task execution efficiency and the overall stability of the system.

[0085] (3) The present invention adopts a multi-round game iterative optimization algorithm. Each cloud computing node dynamically adjusts its resource competition strategy by maximizing its benefit function. In multiple rounds of local games and global collaborative games, the node adjusts its strategy based on feedback information and historical data to gradually approach the optimal resource allocation plan. At the same time, the system continuously optimizes the scheduling strategy through an adaptive learning mechanism and combines task priority and resource load parameters to achieve more accurate resource demand prediction, which greatly improves the accuracy of resource allocation. Compared with the traditional static scheduling method, the present invention improves resource utilization by more than 12%, effectively reducing resource waste and task delays. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] 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:

[0087] Figure 1 This is a flow chart of a method for competitive scheduling of storage, computing and training resources of cloud computing nodes based on an improved model of game theory proposed by the present invention;

[0088] Figure 2 This is a schematic diagram of the adaptive adjustment of the local game model and the global collaborative game model in the competitive scheduling method for storage, computing and training resources of cloud computing nodes based on an improved model of game theory proposed by the present invention. DETAILED DESCRIPTION

[0089] 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.

[0090] refer to Figure 1-2 A method for scheduling storage, computing and training resource competition of cloud computing nodes based on an improved model of game theory includes the following steps:

[0091] S1. In a cloud computing environment, deploy multiple cloud computing nodes distributed in multiple logical partitions, each cloud computing node includes storage resources, computing resources, and training resources, and build a global resource scheduling system;

[0092] S2. Initially divide the cloud computing nodes in each logical partition, collect resource status information of each cloud computing node, and build a local resource competition game model in the logical partition based on the collected resource status information. Each cloud computing node in the logical partition is regarded as an independent game participant, and the nodes formulate resource competition strategies according to their own task requirements, resource load and task priority.

[0093] S3. Execute the local resource competition game model. Through the strategy game process, the cloud computing nodes in the logical partition adjust the resource demand strategy, optimize the use of storage resources, computing resources and training resources, generate local resource competition results, and allocate resources in the logical partition;

[0094] S4. The global resource scheduling system collects the local resource competition results of each logical partition and the remaining available resources in each logical partition. Based on the global collaborative game model in game theory, each logical partition is regarded as a global game participant. Each logical partition adjusts its resource usage strategy according to the local resource competition results and global resource demand.

[0095] S5. Execute the global collaborative game model. Under the global coordination of the global resource scheduling system, each logical partition performs optimal scheduling of global resources through strategy adjustment and information sharing, generates a global resource allocation plan, and coordinates the allocation of storage resources, computing resources, and training resources among the logical partitions.

[0096] S6. The global resource scheduling system collects the resource usage and task execution status of each cloud computing node in real time through the monitoring module, and dynamically updates the resource status information of the cloud computing node according to the dynamic changes of task load, resource bottleneck and global performance of the system;

[0097] S7, the global resource scheduling system adaptively adjusts the resource demand strategies in the local game model and the global collaborative game model, triggers the global resource scheduling system to perform global scheduling optimization when a resource bottleneck or task overload in a logical partition is detected, and prioritizes the allocation of storage resources, computing resources, and training resources from other logical partitions;

[0098] S8. Strategy optimization and learning are performed based on historical game data. The global resource scheduling system stores and analyzes past resource competition results and task execution status. Each cloud computing node updates its resource demand strategy based on the data and makes better resource allocation decisions in future game processes.

[0099] In this embodiment, S1 further includes:

[0100] S11. Construct a physical cluster including multiple logical partitions in a cloud computing environment. Each logical partition is composed of a number of cloud computing nodes including storage resources, computing resources, and training resources. The distribution of each cloud computing node in different logical partitions depends on the network topology and task load requirements of each logical partition.

[0101] S12: Initialize and configure the storage resources, computing resources, and training resources in the cloud computing nodes. i Represents each cloud computing node N i The initial available storage capacity and computing resources Ci Represents each cloud computing node N i The initial available computing power, training resources T i Represents each cloud computing node N i Initial available training capacity;

[0102] S13, according to each cloud computing node N i The geographical location, network bandwidth and task requirements determine the node distribution within each logical partition, so that the storage resources, computing resources and training resources within the logical partition are evenly distributed, forming a global resource scheduling system structure:

[0103] S i +C i +T i =R i ;

[0104] Among them, R i Represents the cloud computing node N i The total resources, S i , C i and T i They represent the initial configuration amounts of storage resources, computing resources, and training resources respectively.

[0105] In this embodiment, S2 further includes:

[0106] S21, initially divide the cloud computing nodes in each logical partition, collect the N ij Resource status information, where N ij represents the i-th cloud computing node in logical partition j;

[0107] S22, collect cloud computing nodes N ij Resource status information, including:

[0108] Storage resource load:

[0109]

[0110] Computing resource load:

[0111]

[0112] Training resource load:

[0113]

[0114] in, Represents the cloud computing nodes N ij of used storage resources, computing resources, and training resources, Represents the cloud computing nodes Nij The total storage resource capacity, total computing resource capacity and total training resource capacity;

[0115] S23. Calculate the cloud computing node N ij Remaining capacity:

[0116] Remaining storage capacity:

[0117]

[0118] Calculate the remaining capacity of resources:

[0119]

[0120] Remaining capacity of training resources:

[0121]

[0122] S24. Obtain cloud computing node N ij The task execution progress P ij , indicating the completion percentage of the current task;

[0123] S25, based on the collected resource status information, construct a local resource competition game model within the logical partition j, and divide each cloud computing node N within the logical partition ij Treated as independent game players;

[0124] S26, cloud computing node N ij According to your own task requirements ij , Resource Load and the task priority U ij Formulate resource competition strategies ij , task requirement D ij Including the required storage resources Computing resources and training resources

[0125] S27. Define cloud computing node N ij The profit function of the local resource competition game model Π ij , used to evaluate the effect of resource competition strategy, the benefit function comprehensively considers resource utilization efficiency, task completion speed and task priority:

[0126]

[0127] Among them, α1, β1, and γ1 are weight coefficients, which reflect the influence of resource demand, task progress, and task priority in the benefit function.

[0128] In this implementation, S3 further includes:

[0129] S31, execute the local resource competition game model to allocate each cloud computing node N in the logical partition j ij As a game participant, each node in the local resource competition game model is based on its current resource demand strategy S ij Participate in resource competition, the goal of the game process is to optimize the resource usage of cloud computing nodes;

[0130] S32, each cloud computing node N ij In the game process, according to its own resource status information and profit function Π ij , through the strategy adjustment process, iteratively optimizes its resource demand strategy S ij , the direction of strategy adjustment is guided by the goal of maximizing the profit function:

[0131]

[0132] Nodes are based on resource load and task requirements D ij Its resource demand strategy S ij Make dynamic adjustments;

[0133] S33, execute multiple rounds of strategy game process, each cloud computing node updates its resource demand strategy S after each round of game. ij , so that the resource utilization efficiency of the node is gradually optimized:

[0134]

[0135] Among them, η1 is the learning rate, represents the resource demand strategy after the tth round of game;

[0136] S34, the node continues to adjust its strategy in the local resource competition game until the preset convergence condition is met. The specific condition is the profit function Π ij The increment is lower than the set threshold ∈:

[0137]

[0138] S35. After the convergence conditions are met, the node generates the final local resource competition result, which includes the final resource demand strategy S of each cloud computing node. ij and the actual allocated storage resources Computing resources and training resources

[0139] S36. Based on the local resource competition results, the cloud computing nodes are allocated resources in the logical partition j. The allocation strategy is based on the final resource demand strategy S of the node. ijAs well as the remaining available resources in the logical partition, the resource requirements of each node can be optimally met.

[0140] In this implementation, S4 further includes:

[0141] S41, the global resource scheduling system collects the local resource competition results of each logical partition j, including each cloud computing node N ij Resource demand strategy S ij and actual allocated storage resources Computing resources and training resources

[0142] S42: The global resource scheduling system collects the remaining available resources in each logical partition, and calculates the remaining available storage resources of each logical partition j based on the dynamic changes of resources and the dependencies between nodes. Computing resources and training resources

[0143]

[0144] in, They represent the remaining storage resources, computing resources, and training resources of the logical partition j at time t, respectively. Represents the cloud computing nodes N ij The total storage resources, total computing resources and total training resource capacity at time t, Represents the cloud computing nodes N ij The storage resources, computing resources, and training resources actually allocated at time t, The cloud computing nodes N ij Dynamic weighting coefficients on storage resources, computing resources, and training resources are used to measure the rate of change of actual node usage resources, ΔT ij (t) represents the cloud computing node N ij The task execution time difference at time t, ΔT max represents the maximum task execution time difference of all nodes in logical partition j, which is used to standardize the time difference, F ij (t) is the dynamic adjustment factor, which represents the cloud computing node N ij Resource utilization adjustment coefficient at time t;

[0145] S43, based on the collected local resource competition results and the remaining available resources of each logical partition, the global resource scheduling system regards each logical partition j as a global game participant to build a global collaborative game model, in which each logical partition j adjusts the global resource demand strategy S according to its local resource competition results and remaining resource conditions j:

[0146]

[0147] Among them, α j and β j The adjustment coefficient for each logical partition reflects the balance between the remaining resources and the demand of each logical partition. The storage, computing and training resource requirements of each node respectively.

[0148] In this implementation, S5 further includes:

[0149] S51, executing the global collaborative game model, with each logical partition j as a global game participant, and the global resource scheduling system summarizes and analyzes the local resource competition results and remaining resource conditions of each logical partition, builds a global resource scheduling strategy, and performs resource sharing and coordination in a global scope;

[0150] S52, the global resource scheduling system uses the global resource demand strategy S of each logical partition j and remaining resources The resource allocation strategy of each logical partition is adjusted through multiple rounds of games. Each logical partition shares its current resource demand and remaining resource status during the game. The game goal is to achieve optimal scheduling of global resources through strategy adjustment.

[0151] S53. Profit function Π of the global cooperative game model j Defined as the balance between resource utilization efficiency and resource supply demand for each logical partition:

[0152]

[0153] Among them, α j1 and β j1 is the balance coefficient, which reflects the degree of attention paid by each logical partition to the remaining resources and task priorities. j represents the overall progress of task completion in logical partition j, They are the storage, computing, and training resource requirements of the node respectively;

[0154] S54, execute multiple rounds of global game, each logical partition adjusts its resource demand strategy based on the feedback information obtained during the global game process j , adjust the direction by maximizing the profit function Π of the global cooperative game model j OK, the specific adjustment strategy is:

[0155]

[0156] Among them, η2 is the learning rate, represents the resource demand strategy after the tth round of global game;

[0157] S55. When the global game model meets the convergence condition: the profit function Π j The increment of is lower than the set threshold ∈, and a global resource allocation plan is generated. The global resource allocation plan determines the final allocation of storage resources, computing resources, and training resources among each logical partition. The allocation result is optimized based on the resource demand and remaining resource situation of each logical partition during the game process.

[0158] In this implementation, S7 further includes:

[0159] S71. The global resource scheduling system monitors the resource usage of each logical partition in real time, collects the load and remaining capacity of the storage resources, computing resources, and training resources of each cloud computing node, and identifies resource bottlenecks or task overloads.

[0160] S72. When the task load in a certain logical partition exceeds the resource threshold, the adaptive adjustment mechanism of the local game model is triggered. Each cloud computing node adjusts its resource competition strategy according to the current resource load, task demand and task priority to cope with the shortage or bottleneck of local resources, and optimizes the resource demand through the iterative process of the local game model.

[0161] S73, the global resource scheduling system performs adaptive adjustment of the global collaborative game model for each logical partition. When a task overload or resource bottleneck is detected, the global game mechanism coordinates the resource flow between multiple logical partitions, and each partition readjusts its global resource demand strategy according to the current local game result;

[0162] S74. When a partition faces resource shortage, the global resource scheduling system prioritizes allocating available storage resources, computing resources, and training resources from other partitions. The partitions coordinate resource flows through global game to maximize global resource utilization efficiency.

[0163] S75, the local game model and the global collaborative game model realize adaptive adjustment in the process, and the global resource scheduling system continuously updates the resource status of each partition and performs dynamic optimization according to the real-time monitoring data, so that the resource usage of the global resource scheduling system is in the optimal state under different task load conditions;

[0164] S76. The global resource scheduling system further optimizes the resource demand strategies of nodes and partitions based on historical game data, and improves the local game model and the global collaborative game model through a continuous adaptive learning mechanism.

[0165] Embodiment 1:

[0166] In order to verify the feasibility and effectiveness of the present invention in practical applications, the experimenters designed an experiment in a large-scale cloud computing environment and selected the public cloud computing node simulation scheduling data sets CloudSim and Google ClusterData. CloudSim is an open source framework widely used to simulate resource scheduling in cloud computing environments. The experimenters generated simulated resource load data through it. The data set includes resource requirements, task execution time and task priority information of different tasks on cloud computing nodes. Google Cluster Data contains real large-scale data center node scheduling information. The data set covers the resource allocation of hundreds of computing nodes and records in detail the resource consumption, allocation strategy and completion time of each task. The experimenters divided the two data sets into training set and test set in a ratio of 7:3.

[0167] In the CloudSim data set, the experimenters simulated a cluster consisting of 1,000 cloud computing nodes. The nodes were distributed in 10 logical partitions. The number of nodes and resource allocation of each logical partition were different. The total storage resources were 100TB, the total computing resources were 5,000 CPU cores, and the total training resources were 3,000 GPU cores. The Google Cluster Data data set contains 250 real computing nodes. The resources of each node include storage, computing and training resources. It records the task data within one week in May 2011. The tasks include computing-intensive and training-intensive tasks.

[0168] Task type and resource requirements: Computation-intensive tasks require a large number of CPU cores and relatively few storage resources. Training-intensive tasks mainly consume GPU cores and storage resources and have medium computing resource requirements.

[0169] Scheduling model setting: In order to verify the effect of the scheduling method based on the improved model of game theory in the present invention, the experimenters first ran the traditional static scheduling algorithm in the CloudSim environment, and then ran the improved resource scheduling algorithm based on game theory in the present invention. In the Google Cluster Data data set, the experimenters used the default scheduling strategy of the Google cluster for comparison.

[0170] During the scheduling process, the experimenters collected the storage, computing, and training resource status of each node, built a local game model, and adjusted each node through multiple rounds of strategies to optimize its resource demand strategy based on task priority and resource load. When the global resource scheduling system detects task overload or resource bottleneck, it uses the global game model to coordinate the flow of resources between logical partitions to ensure efficient completion of the task.

[0171] CloudSim dataset: The simulation task data includes 5,000 tasks. The task execution time ranges from 10 minutes to 2 hours. The resource requirements of each task include storage resources, computing resources, and training resources. The experimenters selected 3,000 tasks as training data and the remaining 2,000 tasks for testing.

[0172] Google Cluster Data: This dataset contains 10 days of task data, totaling 10,000 task records. The task execution time ranges from a few seconds to a few hours. The experimenters selected 7 days of data as the training set and 3 days of data for testing.

[0173] Evaluation indicators: In order to comprehensively evaluate the performance of the method of the present invention, the experimenters used indicators such as task completion time, resource utilization, scheduling response time and resource scheduling success rate:

[0174] Task completion time: The time taken for a task to be submitted and completed, in seconds.

[0175] Resource utilization: The average utilization of various resources (storage, computing, and training), expressed as a percentage.

[0176] Scheduling response time: The time it takes for the system to respond by allocating resources after a task is submitted, measured in seconds.

[0177] Resource scheduling success rate: the ratio of the resources that the scheduling system successfully allocates to tasks during task execution.

[0178] On the CloudSim and Google Cluster Data datasets, the experimenters ran the traditional static scheduling algorithm and the improved game theory model of the present invention respectively. The experimental results are shown in Tables 1 and 2.

[0179] Table 1 Comparison of experimental results on the CloudSim dataset

[0180]

[0181] Table 2 Comparison of experimental results on Google Cluster Data dataset

[0182]

[0183] It can be seen from Tables 1 and 2 above that the task completion time is significantly reduced: in the two data sets of CloudSim and Google ClusterData, the scheduling method of the present invention shortens the task completion time by 33% and 30% respectively compared with the traditional method and Google's default scheduling strategy. This is because the improved game theory model can adaptively adjust in time when resource bottlenecks occur to allocate surplus resources from other logical partitions to the required partitions, thereby avoiding task delays caused by resource competition.

[0184] Significant improvement in resource utilization: Experimental results show that traditional static scheduling methods and Google cluster default strategies perform poorly in terms of resource utilization, with an average resource utilization of about 78%-80%. The improved game theory model of the present invention improves resource utilization to 92%-93%. The improvement is due to the local game model achieving more refined resource allocation within each logical partition, and the global collaborative game model further optimizes resource utilization efficiency through global resource flow.

[0185] Reduction of scheduling response time: When facing high-concurrency tasks, the traditional scheduling method and the default strategy of the Google cluster have long scheduling response times of 1.2 seconds and 1.1 seconds respectively. The scheduling method of the present invention can quickly respond to task requirements, with a scheduling response time of only 0.8 seconds and 0.7 seconds. The game theory model can quickly adapt to changes in node resource load during task peaks through multiple rounds of strategy adjustments, and allocate resources in a timely manner.

[0186] Improvement of resource scheduling success rate: The resource scheduling success rate of the present invention reaches 98% and 97%, which is much higher than the traditional scheduling method and the default strategy of Google cluster. Since the improved model based on game theory can effectively avoid resource competition, it improves the resource matching ability of the system during task execution and ensures that the task can be executed smoothly.

[0187] This embodiment conducts experiments on two data sets, CloudSim and Google Cluster Data, to prove the feasibility and effectiveness of the resource scheduling method of the present invention in a cloud computing environment. By introducing an improved model based on game theory, the present invention can effectively solve resource bottlenecks and overload problems in a multi-task high-concurrency environment, greatly improve resource utilization and scheduling response speed, and ensure that the system can maintain efficient operation under different load conditions.

[0188] The present invention can dynamically optimize storage, computing and training resources within a partition and globally by dividing the resource competition of cloud computing nodes into two layers: local game and global collaborative game model. The local game model is used for resource competition scheduling of nodes within a partition to ensure the optimal allocation of resources within the partition. The global collaborative game model is based on the global resource status and the game results of each partition, and realizes cross-partition resource scheduling through multiple rounds of strategy adjustment. The two-layer game structure breaks through the limitations of traditional static resource scheduling, greatly improves the resource allocation efficiency, avoids resource bottlenecks within a single partition, and reduces the idleness of resources in other partitions to ensure efficient utilization of global resources under load changes.

[0189] The present invention adopts an adaptive adjustment strategy by real-time monitoring of the resource status in the cloud computing environment, so that each node can flexibly adjust the resource competition strategy when the task load changes. When there is a resource bottleneck or task overload, the global scheduling optimization is automatically triggered through the adaptive adjustment model, and available storage, computing and training resources are preferentially allocated from other partitions. This solves the problem of scheduling lag in the prior art, enables the system to quickly adjust the resource allocation plan when facing high-concurrency tasks and dynamic loads, and ensures the task execution efficiency and the overall stability of the system.

[0190] The present invention adopts a multi-round game iterative optimization algorithm. Each cloud computing node dynamically adjusts its resource competition strategy by maximizing its benefit function. In multiple rounds of iterations of local games and global collaborative games, the node adjusts its strategy based on feedback information and historical data to gradually approach the optimal resource allocation plan. At the same time, the system continuously optimizes the scheduling strategy through an adaptive learning mechanism and combines task priority and resource load parameters to achieve more accurate resource demand prediction, which greatly improves the accuracy of resource allocation. Compared with traditional static scheduling methods, the present invention improves resource utilization by more than 12%, effectively reducing resource waste and task delays.

[0191] 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 scheduling storage and computing resource competition of cloud computing nodes based on an improved model of game theory, characterized in that: The steps include: S1. In a cloud computing environment, deploy multiple cloud computing nodes distributed in multiple logical partitions, each cloud computing node includes storage resources, computing resources, and training resources, and build a global resource scheduling system; S2. Initially divide the cloud computing nodes in each logical partition, collect resource status information of each cloud computing node, and build a local resource competition game model in the logical partition based on the collected resource status information. Each cloud computing node in the logical partition is regarded as an independent game participant, and the nodes formulate resource competition strategies according to their own task requirements, resource load and task priority. S3. Execute the local resource competition game model. Through the strategy game process, the cloud computing nodes in the logical partition adjust the resource demand strategy, optimize the use of storage resources, computing resources and training resources, generate local resource competition results, and allocate resources in the logical partition; S4. The global resource scheduling system collects the local resource competition results of each logical partition and the remaining available resources in each logical partition. Based on the global collaborative game model in game theory, each logical partition is regarded as a global game participant. Each logical partition adjusts its resource usage strategy according to the local resource competition results and global resource demand. S5. Execute the global collaborative game model. Under the global coordination of the global resource scheduling system, each logical partition performs optimal scheduling of global resources through strategy adjustment and information sharing, generates a global resource allocation plan, and coordinates the allocation of storage resources, computing resources, and training resources among the logical partitions. S6. The global resource scheduling system collects the resource usage and task execution status of each cloud computing node in real time through the monitoring module, and dynamically updates the resource status information of the cloud computing node according to the dynamic changes of task load, resource bottleneck and global performance of the system; S7, the global resource scheduling system adaptively adjusts the resource demand strategies in the local game model and the global collaborative game model, triggers the global resource scheduling system to perform global scheduling optimization when a resource bottleneck or task overload in a logical partition is detected, and prioritizes the allocation of storage resources, computing resources, and training resources from other logical partitions; S8. Strategy optimization and learning are performed based on historical game data. The global resource scheduling system stores and analyzes past resource competition results and task execution status. Each cloud computing node updates its resource demand strategy based on the data and makes better resource allocation decisions in future game processes.

2. According to the improved model of game theory, the method for scheduling storage and computing resources competition of cloud computing nodes is characterized in that: The S1 further comprises: S11. Construct a physical cluster including multiple logical partitions in a cloud computing environment. Each logical partition is composed of a number of cloud computing nodes including storage resources, computing resources, and training resources. The distribution of each cloud computing node in different logical partitions depends on the network topology and task load requirements of each logical partition. S12: Initialize and configure the storage resources, computing resources, and training resources in the cloud computing nodes. i Represents each cloud computing node N i The initial available storage capacity and computing resources C i Represents each cloud computing node N i The initial available computing power, training resources T i Represents each cloud computing node N i Initial available training capacity; S13, according to each cloud computing node N i The geographical location, network bandwidth and task requirements determine the node distribution within each logical partition, so that the storage resources, computing resources and training resources within the logical partition are evenly distributed, forming a global resource scheduling system structure: S i +C i +T i =R i ; Among them, R i Represents the cloud computing node N i The total resources, S i , C i and T i They represent the initial configuration amounts of storage resources, computing resources, and training resources respectively.

3. According to the method of claim 1, the method is characterized in that: The S2 further comprises: S21, initially divide the cloud computing nodes in each logical partition, collect the N ij Resource status information, where N ij represents the i-th cloud computing node in logical partition j; S22, collect cloud computing nodes N ij Resource status information, including: Storage resource load: Computing resource load: Training resource load: in, Represents the cloud computing nodes N ij of used storage resources, computing resources, and training resources, Represents the cloud computing nodes N ij The total storage resource capacity, total computing resource capacity and total training resource capacity; S23. Calculate the cloud computing node N ij Remaining capacity: Remaining storage capacity: Calculate the remaining capacity of resources: Remaining capacity of training resources: S24. Obtain cloud computing node N ij The task execution progress P ij , indicating the completion percentage of the current task; S25, based on the collected resource status information, construct a local resource competition game model within the logical partition j, and divide each cloud computing node N within the logical partition ij Treated as independent game players; S26, cloud computing node N ij According to your own task requirements ij , Resource Load and the task priority U ij Formulate resource competition strategies ij , task requirement D ij Including the required storage resources Computing resources and training resources S27. Define cloud computing node N ij The profit function of the local resource competition game model Π ij , used to evaluate the effect of resource competition strategy, the benefit function comprehensively considers resource utilization efficiency, task completion speed and task priority: Among them, α1, β1, and γ1 are weight coefficients, which reflect the influence of resource demand, task progress, and task priority in the benefit function.

4. According to the method of claim 1, the method for scheduling storage and computing resources competition of cloud computing nodes based on the improved model of game theory is characterized in that: The S3 further comprises: S31, execute the local resource competition game model to allocate each cloud computing node N in the logical partition j ij As a game participant, each node in the local resource competition game model is based on its current resource demand strategy S ij Participate in resource competition, the goal of the game process is to optimize the resource usage of cloud computing nodes; S32, each cloud computing node N ij In the game process, according to its own resource status information and profit function Π ij , through the strategy adjustment process, iteratively optimizes its resource demand strategy S ij , the direction of strategy adjustment is guided by the goal of maximizing the profit function: Nodes are based on resource load and task requirements D ij Its resource demand strategy S ij Make dynamic adjustments; S33, execute multiple rounds of strategy game process, each cloud computing node updates its resource demand strategy S after each round of game. ij , so that the resource utilization efficiency of the node is gradually optimized: Among them, η1 is the learning rate, represents the resource demand strategy after the tth round of game; S34, the node continues to adjust its strategy in the local resource competition game until the preset convergence condition is met. The specific condition is the profit function Π ij The increment is lower than the set threshold ∈: S35. After the convergence conditions are met, the node generates the final local resource competition result, which includes the final resource demand strategy S of each cloud computing node. ij and the actual allocated storage resources Computing resources and training resources S36. Based on the local resource competition results, the cloud computing nodes are allocated resources in the logical partition j. The allocation strategy is based on the final resource demand strategy S of the node. ij As well as the remaining available resources in the logical partition, the resource requirements of each node can be optimally met.

5. According to the method of claim 1, the method for scheduling storage and computing resources competition of cloud computing nodes based on the improved model of game theory is characterized in that: The S4 further comprises: S41, the global resource scheduling system collects the local resource competition results of each logical partition j, including each cloud computing node N ij Resource requirement strategy S ij and actual allocated storage resources Computing resources and training resources S42: The global resource scheduling system collects the remaining available resources in each logical partition, and calculates the remaining available storage resources of each logical partition j based on the dynamic changes of resources and the dependencies between nodes. Computing resources and training resources in, They represent the remaining storage resources, computing resources, and training resources of the logical partition j at time t, respectively. Represents the cloud computing nodes N ij The total storage resources, total computing resources and total training resource capacity at time t, Represents the cloud computing nodes N ij The storage resources, computing resources, and training resources actually allocated at time t, The cloud computing nodes N ij Dynamic weighting coefficients on storage resources, computing resources, and training resources are used to measure the rate of change of actual node usage resources, ΔT ij (t) represents the cloud computing node N ij The task execution time difference at time t, ΔT max represents the maximum task execution time difference of all nodes in logical partition j, which is used to standardize the time difference, F ij (t) is the dynamic adjustment factor, which represents the cloud computing node N ij Resource utilization adjustment coefficient at time t; S43, based on the collected local resource competition results and the remaining available resources of each logical partition, the global resource scheduling system regards each logical partition j as a global game participant to build a global collaborative game model, in which each logical partition j adjusts the global resource demand strategy S according to its local resource competition results and remaining resource conditions j : Among them, α j and β j The adjustment coefficient for each logical partition reflects the balance between the remaining resources and the demand of each logical partition. The storage, computing and training resource requirements of each node respectively.

6. According to the method of claim 1, the method for scheduling storage and computing resources competition of cloud computing nodes based on the improved model of game theory is characterized in that: The S5 further comprises: S51, executing the global collaborative game model, with each logical partition j as a global game participant, and the global resource scheduling system summarizes and analyzes the local resource competition results and remaining resource conditions of each logical partition, builds a global resource scheduling strategy, and performs resource sharing and coordination in a global scope; S52, the global resource scheduling system uses the global resource demand strategy S of each logical partition j and remaining resources The resource allocation strategy of each logical partition is adjusted through multiple rounds of games. Each logical partition shares its current resource demand and remaining resource status during the game. The game goal is to achieve optimal scheduling of global resources through strategy adjustment. S53. Profit function Π of the global cooperative game model j Defined as the balance between resource utilization efficiency and resource supply demand for each logical partition: Among them, α j1 and β j1 is the balance coefficient, which reflects the degree of attention paid by each logical partition to the remaining resources and task priorities. j represents the overall progress of task completion in logical partition j, They are the storage, computing, and training resource requirements of the node respectively; S54, execute multiple rounds of global game, each logical partition adjusts its resource demand strategy based on the feedback information obtained during the global game process j , adjust the direction by maximizing the profit function Π of the global cooperative game model j OK, the specific adjustment strategy is: Among them, η2 is the learning rate, represents the resource demand strategy after the tth round of global game; S55. When the global game model meets the convergence condition: the profit function Π j The increment of is lower than the set threshold ∈, and a global resource allocation plan is generated. The global resource allocation plan determines the final allocation of storage resources, computing resources, and training resources among each logical partition. The allocation result is optimized based on the resource demand and remaining resource situation of each logical partition during the game process.

7. According to claim 1, a method for scheduling storage and computing resources competition of cloud computing nodes based on an improved model of game theory, characterized in that: The S7 further comprises: S71. The global resource scheduling system monitors the resource usage of each logical partition in real time, collects the load and remaining capacity of the storage resources, computing resources, and training resources of each cloud computing node, and identifies resource bottlenecks or task overloads. S72. When the task load in a certain logical partition exceeds the resource threshold, the adaptive adjustment mechanism of the local game model is triggered. Each cloud computing node adjusts its resource competition strategy according to the current resource load, task demand and task priority to cope with the shortage or bottleneck of local resources, and optimizes the resource demand through the iterative process of the local game model. S73, the global resource scheduling system performs adaptive adjustment of the global collaborative game model for each logical partition. When a task overload or resource bottleneck is detected, the global game mechanism coordinates the resource flow between multiple logical partitions, and each partition readjusts its global resource demand strategy according to the current local game result; S74. When a partition faces resource shortage, the global resource scheduling system prioritizes allocating available storage resources, computing resources, and training resources from other partitions. The partitions coordinate resource flows through global game to maximize global resource utilization efficiency. S75, the local game model and the global collaborative game model realize adaptive adjustment in the process, and the global resource scheduling system continuously updates the resource status of each partition and performs dynamic optimization according to the real-time monitoring data, so that the resource usage of the global resource scheduling system is in the optimal state under different task load conditions; S76. The global resource scheduling system further optimizes the resource demand strategies of nodes and partitions based on historical game data, and improves the local game model and the global collaborative game model through a continuous adaptive learning mechanism.

Images