International data center resource scheduling method and system based on deep reinforcement learning

Abstract

The present invention relates to the technical field of resource scheduling, and more specifically, to an international data center resource scheduling method and system based on deep reinforcement learning, including: obtaining a resource state matrix of an international data center; obtaining a task demand matrix; obtaining an energy consumption matrix, wherein the energy consumption matrix includes the unit resource energy consumption of each data center; constructing a resource scheduling benefit function based on the resource state matrix and the task demand matrix; optimizing a resource allocation strategy according to the energy consumption matrix and the resource scheduling benefit function, and generating an optimized resource allocation matrix; determining a task migration path based on the optimized resource allocation matrix, and generating a task migration matrix; outputting the optimized resource allocation matrix and the task migration matrix for realizing resource optimization and task scheduling of the international data center, wherein the optimized output of the resource allocation matrix directly affects the input of network topology adjustment, and forms a highly coordinated optimization system.

Description

Technical Field

[0001] The present invention relates to the technical field of resource scheduling, and more specifically, to an international data center resource scheduling method and system based on deep reinforcement learning. Background Art

[0002] With the rapid development of cloud computing, big data and artificial intelligence technologies, international data centers have become important infrastructure to support global computing and storage needs. The operating efficiency, energy consumption management and task scheduling strategies of data centers directly affect their service capabilities and operating costs. However, in practical applications, resource scheduling faces highly dynamic, diverse and complex problems. Traditional resource allocation and task scheduling methods have many technical bottlenecks in dealing with these challenges.

[0003] In the prior art, resource scheduling methods for data centers are mainly based on predefined static rules or simple heuristic algorithms. These methods usually rely on fixed resource allocation strategies to allocate tasks to specific servers or nodes, and cannot dynamically adapt to real-time changes in load. For example, traditional methods may statically match the computing requirements of tasks with the available resources in data centers, ignoring the multiple requirements of tasks for bandwidth, latency, and energy consumption. Although this allocation method is simple to implement, it exposes the following problems in actual operation:

[0004] Static rules lack the ability to respond to dynamic changes in task loads, often resulting in some resource nodes being overloaded while other node resources are idle. For example, real-time tasks may be delayed due to local overload, while non-real-time tasks may occupy key resources, resulting in low overall resource utilization.

[0005] Existing technologies lack global consideration of path optimization during task migration, and usually make decisions based only on the shortest path or the lowest energy consumption. This approach easily leads to increased migration latency, especially in the case of complex network topology or limited bandwidth, and task migration efficiency is difficult to guarantee.

[0006] Traditional methods fail to effectively combine resource allocation and energy optimization, and only run tasks within a fixed resource range, failing to comprehensively consider load balancing and energy efficiency. Overloaded operation or idle resources in a data center will significantly increase the overall energy consumption of the system, thereby increasing operating costs.

[0007] Traditional methods are usually based on simple heuristic rules or linear optimization models, and their algorithms show low adaptability in complex environments. For example, when the priority of tasks, load dynamics, and network topology change, these algorithms cannot adjust their strategies in time, resulting in reduced resource scheduling efficiency and stability.

[0008] The existing technology has the following main problems when facing dynamic resource scheduling requirements:

[0009] 1. Lack of dynamic response capability: Existing methods have difficulty in real-time perception and prediction of changes in data center resource status and task requirements. For example, when the task load fluctuates greatly, traditional methods cannot dynamically adjust the allocation plan, resulting in a decline in service quality.

[0010] 2. Lack of global optimization perspective: Traditional task migration path optimization only considers a single factor (such as the shortest path), but fails to integrate multiple factors such as network latency, bandwidth utilization, and energy consumption, making it difficult to achieve global optimization.

[0011] 3. Insufficient coordination between resources and networks: Existing methods fail to achieve the coordinated optimization of resource allocation and network topology adjustment, and ignore the mutual influence between task migration and resource utilization. For example, task migration may lead to new resource overload or network bottleneck, which is not fully considered.

[0012] 4. Algorithm limitations: Heuristic rules and linear optimization algorithms show significant limitations in complex, multi-objective environments. For example, in the task scheduling process, resolving resource conflicts and improving energy efficiency often conflict with each other, and traditional methods are difficult to take both into account. Summary of the invention

[0013] In view of the shortcomings of the prior art, the present invention proposes an international data center resource scheduling method and system based on deep reinforcement learning. Through dynamically optimized resource allocation strategies and task migration paths, the present invention aims to solve the following technical problems: improve the resource utilization efficiency of the data center; shorten the task completion time to meet real-time requirements; reduce system energy consumption to achieve green computing; optimize the task migration path to reduce migration costs; achieve coordinated optimization of resource allocation and network topology to comprehensively improve scheduling performance.

[0014] The present invention provides an international data center resource scheduling method based on deep reinforcement learning, comprising:

[0015] The acquisition steps include:

[0016] Obtain a resource status matrix of the international data center, where the resource status matrix includes resource utilization, computing capacity, and network load status of each data center;

[0017] Obtain a task requirement matrix, where the task requirement matrix includes the computing requirement, storage requirement, and latency requirement of the task;

[0018] Obtain an energy consumption matrix, where the energy consumption matrix includes unit resource energy consumption of each data center;

[0019] Processing steps include:

[0020] Constructing a resource scheduling benefit function based on the resource status matrix and the task requirement matrix;

[0021] Optimizing the resource allocation strategy according to the energy consumption matrix and the resource scheduling benefit function, and generating an optimized resource allocation matrix;

[0022] Based on the optimized resource allocation matrix, determine the task migration path and generate a task migration matrix;

[0023] Output steps include:

[0024] The optimized resource allocation matrix and the task migration matrix are outputted to realize resource optimization and task scheduling of the international data center.

[0025] Preferably, the obtaining step further comprises:

[0026] Obtain a network delay matrix, where the network delay matrix includes network transmission delays between data centers;

[0027] The network delay matrix is ​​used as an input for optimizing resource allocation strategy to comprehensively consider the network delay cost of task migration.

[0028] Preferably, the processing steps specifically include:

[0029] Based on the resource status matrix and task requirement matrix, a resource scheduling benefit function F is constructed:

[0030]

[0031] Where: F is the resource scheduling benefit value; w ij is a weight parameter, indicating the importance of task j in data center i; r ij is the resource utilization rate, which indicates the resource consumption ratio of task j in data center i.

[0032] Preferably, the optimized resource allocation strategy specifically includes:

[0033] Minimize resource energy consumption by optimizing the objective function J:

[0034]

[0035] Where: E i is the unit resource energy consumption of the ith data center, C i is the total computing capacity of the ith data center; λ is the regularization coefficient, which controls the scheduling variation; is the Frobenius norm, which is used to limit the difference between the optimized resource matrix and the original matrix.

[0036] Preferably, the optimization of the task migration path is based on the following objective function:

[0037]

[0038] Where: P is the task migration path matrix; p ij is the number of migration tasks from data center i to j; c ij The cost of task migration includes network latency and energy consumption.

[0039] Preferably, the task migration path further satisfies the following constraints:

[0040]

[0041] It indicates that the inflow and outflow of task migration are balanced.

[0042] Preferably, the optimized resource allocation matrix is ​​further combined with the network topology adjustment strategy, using the following comprehensive optimization objective function:

[0043]

[0044] Among them, Z is the comprehensive optimization benefit, and μ is the weight coefficient of task migration cost.

[0045] Preferably, the network topology adjustment strategy is optimized based on an integer linear programming problem, wherein the optimization objectives are to minimize network delay and maximize bandwidth utilization.

[0046] Preferably, the method further includes an exception handling step, in which when the resource utilization of any data center exceeds a preset threshold, the task migration path is dynamically adjusted and the resource allocation matrix is ​​recalculated.

[0047] The international data center resource scheduling system based on deep reinforcement learning for executing the method comprises:

[0048] Data collection module, used to collect the resource status matrix, task demand matrix and energy consumption matrix of the international data center;

[0049] Data processing module, used to construct resource scheduling benefit function and optimize resource allocation strategy based on collected data;

[0050] The scheduling module is used to determine the task migration path based on the optimized resource allocation strategy;

[0051] The output module is used to output the optimized resource allocation matrix and task migration path.

[0052] The method of the present invention establishes a collaborative optimization model of dynamic resource scheduling and task migration by introducing deep reinforcement learning technology, which is specifically embodied in the following aspects:

[0053] 1. Synergy between dynamic adaptability and global optimization: By introducing the policy optimization mechanism of deep reinforcement learning, the agent can perceive the resource status and task requirements in real time and dynamically adjust the allocation strategy. At the same time, the reward function combines multiple objective factors such as resource utilization, migration cost and energy consumption to ensure the global nature of policy optimization.

[0054] 2. Complementarity and superposition of resource allocation and network topology: The method of the present invention solves the problem of independence between resources and networks through the joint optimization of the resource allocation matrix and the network topology adjustment strategy. Specifically, the optimization output of the resource allocation matrix directly affects the input of the network topology adjustment, forming a highly coordinated optimization system.

[0055] 3. Solution to the technical contradiction between migration path and energy efficiency optimization: The optimization goal of task migration path takes into account both migration cost and energy efficiency optimization, and achieves a balance between the two by introducing dynamic weight parameters. For example, in scenarios with high real-time requirements, low-latency paths are preferred; and in scenarios with strict energy consumption constraints, low-energy consumption paths are preferred.

[0056] The beneficial effects of the present invention are as follows:

[0057] 1. Significantly improved resource utilization: Through the dynamic adjustment mechanism of reinforcement learning, the remaining resources of the data center can be fully utilized, resource waste can be reduced, and resource utilization can be significantly improved.

[0058] 2. The task completion time is significantly shortened: Through the migration path and allocation strategy optimized by deep reinforcement learning, the method of the present invention significantly shortens the task completion time in real-time task scenarios and meets the service requirements of high-priority tasks.

[0059] 3. Significant reduction in system energy consumption: By optimizing the dynamic constraints of the objective function on energy consumption, the method of the present invention achieves energy efficiency optimization during the operation of the data center and reduces the overall system energy consumption.

[0060] 4. Greatly improved migration efficiency: Through the optimized design of the task migration path, the present invention effectively reduces the delay and cost of task migration and improves the migration efficiency.

[0061] 5. Collaborative optimization of global performance: The joint optimization of resource allocation, task migration, and network topology adjustment has comprehensively improved the overall performance of the data center, making it more adaptable and robust.

[0062] In summary, the method and system of the present invention solve the shortcomings of the existing technology through deep reinforcement learning technology, provide an efficient, intelligent and energy-saving solution for international data center resource scheduling, and show superiority in dynamic task environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is the overall framework of the method of the present invention.

[0064] Figure 2 It is a logic block diagram of the data acquisition module of the present invention.

[0065] Figure 3 It is a logic block diagram of the data processing module of the present invention.

[0066] Figure 4 It is a logic block diagram of the scheduling module of the present invention.

[0067] Figure 5 It is a logic block diagram of the output module of the present invention. DETAILED DESCRIPTION

[0068] Please refer to Figure 1-5 The present invention provides an international data center resource scheduling method based on deep reinforcement learning, including: obtaining a resource status matrix of the international data center, wherein the resource status matrix includes the resource utilization rate, computing power and network load status of each data center; obtaining a task requirement matrix, wherein the task requirement matrix includes the computing requirement, storage requirement and latency requirement of the task; obtaining an energy consumption matrix, wherein the energy consumption matrix includes the unit resource energy consumption of each data center.

[0069] Based on the resource status matrix and the task demand matrix, a resource scheduling benefit function is constructed; based on the energy consumption matrix and the resource scheduling benefit function, a resource allocation strategy is optimized to generate an optimized resource allocation matrix; based on the optimized resource allocation matrix, a task migration path is determined to generate a task migration matrix.

[0070] The optimized resource allocation matrix and the task migration matrix are outputted to realize resource optimization and task scheduling of the international data center.

[0071] In a preferred embodiment of the present invention, the resource status matrix can be obtained in the following manner:

[0072] Assume that there are n data centers in the system, and each data center has m types of resources (such as CPU, GPU, storage, etc.). The specific form of the resource state matrix R is:

[0073]

[0074] Where: R is the resource status matrix; r ij is the utilization rate of the jth resource in the ith data center, ranging from [0,1]; n is the number of data centers; and m is the number of resource types.

[0075] The resource status matrix R is an n×m matrix, where n represents the number of data centers and m represents the number of resource types. Each element r of the matrix ij Indicates the utilization rate of the jth resource in the i-th data center, with a value range of [0,1], where 0 indicates that the resource is completely unused and 1 indicates that the resource is fully occupied. By representing the resource status in matrix form, linear transformation and matrix operations can be easily performed, which is convenient for subsequent optimization algorithm processing.

[0076] In the international data center, the resource status matrix R is used to describe the resource usage of each data center. Through this matrix, we can intuitively understand the resource utilization of each data center, thus providing basic data for subsequent resource scheduling. For example, if the CPU utilization of a data center is high, while the storage resources of another data center are relatively abundant, these differences can be quickly identified through the matrix R, and then a reasonable scheduling strategy can be formulated.

[0077] For example, if the CPU usage of data center A is 60% and the storage usage is 40%, the corresponding r ij 0.6 and 0.4 respectively.

[0078] The task demand matrix D represents the demand of the current task for each resource:

[0079]

[0080] Among them, d ij is the demand of the i-th task for the j-th resource, and the unit is the corresponding resource amount (such as the number of CPU cores or storage capacity). Through this matrix representation, the resource status and task requirements of the data center can be clearly described, providing a complete input for the subsequent benefit function construction.

[0081] The energy consumption matrix E represents the energy consumption per unit resource of each data center:

[0082]

[0083] Among them, e ij The unit energy consumption of the i-th data center using the j-th resource type, in watts per unit of resource.

[0084] By obtaining these matrices, we can ensure that resource availability and energy consumption characteristics are taken into account when allocating tasks.

[0085] In a preferred embodiment of the present invention, firstly, a resource scheduling benefit function F is constructed based on the resource state matrix and the task demand matrix:

[0086]

[0087] Among them, F is the resource scheduling benefit value, which indicates the pros and cons of the current allocation strategy; w ij is a weight parameter, which reflects the importance of the task to the jth type of resources. For example, tasks with higher real-time requirements have a greater computing resource weight; r ij is the utilization rate of the j-th resource in the ith data center.

[0088] The resource scheduling benefit function F is obtained by weighted summing the resource state matrix R and the weight matrix W. Each element w of the weight matrix W ij Indicates the weight of the j-th resource in the i-th data center, reflecting the importance and priority of different resources. ij represents the remaining utilization potential of the jth resource in the i-th data center, that is, the unused part of the resource. By maximizing F, we can ensure that resource allocation is more reasonable and maximize the overall benefit. If the weight matrix W is non-negative and the resource utilization rate r ij is also limited to [0,1], then F is a convex function and the global optimal solution can be found through a convex optimization algorithm.

[0089] The resource scheduling benefit function F is used to quantify the effectiveness of the resource allocation strategy. By introducing the weight matrix W, the impact of different resources on the total benefit can be adjusted according to their importance. For example, some critical tasks may require more CPU resources, while other tasks may rely more on storage space. Therefore, by adjusting the weight matrix W, it is possible to ensure that resource allocation is more reasonable and maximize the overall benefit.

[0090] In practical applications, w ij For example, if the computing task is highly dependent on the CPU, the CPU weight w ij It can be set to 0.8, and the storage resource weight is set to 0.2.

[0091] The benefit value of each allocation scheme is calculated through this function, and the resource allocation strategy is further optimized in combination with the energy consumption matrix.

[0092] In a preferred embodiment of the present invention, the optimized resource allocation matrix R' and the task migration matrix P are outputs of the method. Specifically, R' reflects the optimized resource usage status of each data center; P represents the specific scheme of migrating tasks from one data center to another.

[0093] Optimization strategy iteration:

[0094]

[0095] First, calculate the gradient of the objective function:

[0096]

[0097] Then use gradient descent to update R′:

[0098]

[0099] Objective Function In the above example, the first term F(R′,W) is the resource scheduling benefit function, and the second term is a regularization term used to control the magnitude of scheduling changes. The regularization term is calculated by the Frobenius norm. Measure the difference between the new and old resource states to prevent the scheduling changes from being too drastic and causing system instability. Gradient descent is a commonly used optimization algorithm that gradually approaches the optimal solution by calculating the gradient of the objective function and gradually adjusting the parameters. It indicates the direction in which the current resource state matrix R′ should be adjusted to maximize the objective function. The learning rate η controls the step size of each iteration and affects the optimization speed and stability.

[0100] In practical applications, resource scheduling is not only about maximizing resource utilization, but also about considering the smoothness of scheduling changes and the stability of the system. It can optimize resource allocation while avoiding frequent or drastic scheduling changes, thereby reducing the impact on system performance. The gradient descent method uses steady steps to approach the optimal solution, ensuring that the resource allocation plan is continuously optimized.

[0101] Input the resource status matrix R and the resource scheduling benefit function F, and output the optimized resource status matrix R′.

[0102] Task migration optimization based on optimal path

[0103]

[0104] The construction of the task transfer matrix depends on the following constraints:

[0105]

[0106] That is, the inflow and outflow of tasks need to be balanced.

[0107] The goal of task migration optimization is to find a task allocation scheme P that minimizes the total migration cost. The migration cost is given by the migration cost matrix C, where c ij represents the cost of migrating a unit task from data center i to data center j, including factors such as network bandwidth and latency. The task migration optimization problem can be regarded as a linear programming problem with the constraints Ensure that the inflow and outflow of tasks are balanced to avoid resource waste or overload. By solving linear programming problems, the optimal task migration solution can be found. In practical applications, the shortest path algorithm (such as Dijkstra algorithm or A * algorithm) to find the optimal migration path from one data center to another, further reducing the migration cost.

[0108] In the international data center, task migration is an important part of resource scheduling. By minimizing the migration cost, network bandwidth consumption and latency can be reduced, and the overall performance of the system can be improved. Constraints ensure that the inflow and outflow of tasks are balanced to avoid resource waste or overload. The optimal path algorithm is used to find the most economical task migration solution to ensure the efficiency of resource scheduling.

[0109] Input the resource status matrix R′ and output the optimized resource status matrix P.

[0110] The optimized results can significantly improve the resource utilization of task scheduling while reducing the overall energy consumption of the data center.

[0111] In a preferred embodiment of the present invention, the acquisition step further includes: acquiring a network delay matrix, wherein the network delay matrix includes the network transmission delay between each data center; using the network delay matrix as an input for optimizing the resource allocation strategy to comprehensively consider the network delay cost of task migration.

[0112] The network delay matrix L is used to represent the network transmission delay between data centers:

[0113]

[0114] Among them: ij is the transmission delay from data center i to data center j, in milliseconds (ms); l ii =0, indicating that there is no delay in the local data center. For example, assuming that the transmission delay between data centers A and B is 10 milliseconds, then l AB =10.

[0115] In a preferred embodiment of the present invention, the processing step specifically includes: constructing a resource scheduling benefit function F based on the resource status matrix and the task demand matrix:

[0116]

[0117] By constructing the resource scheduling benefit function F, we can clearly evaluate the pros and cons of the current resource allocation and provide a quantitative reference for subsequent optimization steps. ij When it is small, it indicates that the j-th resource in the i-th data center still has a large utilization space and can be assigned tasks first.ij This further ensures that the importance of the task is accurately quantified.

[0118] In summary, the present invention systematically optimizes the resource scheduling of the international data center through the above steps, and has the beneficial effects of high efficiency and low energy consumption.

[0119] In a preferred embodiment of the present invention, the optimization resource allocation strategy specifically includes: minimizing resource energy consumption by optimizing the objective function J:

[0120]

[0121] Where, is the unit resource energy consumption of the ith data center; r ij is the utilization rate of the jth resource in the ith data center; C i is the total computing capacity of the ith data center; λ is the regularization coefficient, which is used to control the amplitude of scheduling adjustment; is the Frobenius norm, which is used to constrain the difference between the optimized resource allocation matrix R′ and the original matrix R.

[0122] In a preferred embodiment of the present invention, the first term of the objective function J is used to minimize the overall resource energy consumption of the data center. Assume that the CPU unit energy consumption E of data center A is A is 0.5 watts per core, and the GPU unit energy consumption is 1.0 watts per core. If the CPU current utilization rate is r A,CPU =0.6, computing power C A =100 cores, then the resource consumption can be calculated as:

[0123]

[0124] The second term controls the adjustment range of the resource allocation matrix by introducing a regularization parameter λ, thereby avoiding network delays and migration costs caused by excessive scheduling changes. Preferably, in an embodiment of the present invention, the value range of the regularization coefficient λ is 0.01 to 0.1, and is adjusted according to the real-time requirements of the scheduling task. Through the constraints of the above optimization objective function, a balance can be achieved between energy consumption optimization and scheduling stability.

[0125] Through the constraints of the above optimization objective function, a balance can be achieved between energy consumption optimization and scheduling stability.

[0126] In a preferred embodiment of the present invention, the optimization of the task migration path is based on the following objective function:

[0127]

[0128] Among them, P is the task migration path matrix; p ijis the number of migration tasks from data center i to data center j; c ij The cost of task migration includes network latency and energy consumption.

[0129] Preferably, in one embodiment of the present invention, the optimization of the task migration path matrix P needs to comprehensively consider the network delay and energy consumption cost between data centers. ij It can be provided by the network delay matrix L. For example, if the delay between data centers A and B is 20 milliseconds and the energy consumption of a single task migration is 10 watts, the total cost is:

[0130] c AB =20+10=30

[0131] Furthermore, in order to ensure the load balance of the migration task, the method of the present invention adopts the following constraints:

[0132]

[0133] This constraint means that the inflow of tasks in each data center must be equal to the outflow, thus avoiding the problem of biased resource usage. For example, when the total task migration is 100 units, and the number of tasks flowing from data center A to data center B is 60, the number of tasks returning from B to A must be 40. By following the above optimization process, the migration cost can be significantly reduced and the overall resource scheduling efficiency can be improved.

[0134] In the preferred embodiment of the present invention, the above constraints are specially introduced to ensure the balance of the task migration path. Task flow balance means that the task inflow and outflow of any data center need to be consistent to prevent a single data center from affecting the overall stability of the system due to excessive load. For example, if data center A has the following task migration volume:

[0135] Outflow: p A→B =50,p A→C =30

[0136] The inflow volume must satisfy:

[0137] p B→A +p C→A =80

[0138] This balance condition can not only improve the resource utilization of the data center, but also effectively reduce the delay and migration energy consumption during the scheduling process.

[0139] In a preferred embodiment of the present invention, the optimized resource allocation matrix is ​​further combined with the network topology adjustment strategy, using the following comprehensive optimization objective function:

[0140]

[0141] Among them: Z is the comprehensive optimization benefit, w ij is the weight parameter of task importance; r ij is the resource utilization rate; p ij is the task migration amount; c ij is the migration cost; μ is a trade-off parameter used to adjust the balance between resource utilization benefits and migration costs.

[0142] The comprehensive optimization goal Z is a multi-objective optimization problem that takes into account both resource utilization and migration cost. Represents the maximization of resource utilization, the second term Indicates the minimization of migration cost. By introducing the trade-off parameter μ, the optimization goal can be flexibly adjusted in different application scenarios to ensure that resource scheduling is both efficient and economical. In multi-objective optimization problems, there is usually no single optimal solution, but a set of Pareto optimal solutions. By adjusting the trade-off parameter μ, a trade-off can be made between different Pareto optimal solutions to find the scheduling solution that best suits the current scenario.

[0143] The comprehensive optimization target Z is the core of the entire scheduling process. It not only takes into account the resource utilization, but also the cost of task migration. By introducing the trade-off parameter μ, the optimization target can be flexibly adjusted in different application scenarios to ensure that resource scheduling is both efficient and economical. For example, under high load conditions, the value of μ can be appropriately increased to give priority to reducing migration costs; while under low load conditions, the value of μ can be reduced to give priority to improving resource utilization.

[0144] Input the optimized resource state matrix R′ and task migration matrix P, and output the global optimal resource scheduling solution π * .

[0145] In a preferred embodiment of the present invention, the comprehensive optimization objective function comprehensively improves the efficiency of resource scheduling by combining resource allocation and task migration.

[0146] Functions of resource utilization part:

[0147]

[0148] Used to maximize the potential utilization of resources. For example, if the task has a high weight on the GPU ij =0.9, and the current GPU usage rate r ij =0.3, the task will be assigned to GPU first.

[0149] Function of the migration cost part:

[0150]

[0151] Preferably, in the method of the present invention, the value range of the trade-off parameter μ is 0.1 to 1.0, and when the real-time requirement is high, μ is reduced to give priority to reducing the migration delay.

[0152] In summary, through the constraints and adjustments of the above-mentioned comprehensive optimization objective function, the method of the present invention can achieve global optimization of the international data center and provide an efficient and low-cost resource scheduling solution for practical application scenarios.

[0153] The network topology adjustment strategy is optimized based on an integer linear programming problem, wherein the optimization objectives are to minimize network delay and maximize bandwidth utilization.

[0154] In a preferred embodiment of the present invention, in order to further optimize the overall performance of resource scheduling, the network topology adjustment strategy is optimized by the following integer linear programming problem. The objective function is set as follows:

[0155]

[0156] Where T is the optimization target value, which represents the overall performance after the network topology is adjusted; ij is the network delay between data center i and data center j, in milliseconds; b ij is the network bandwidth between data center i and data center j, in Mbps; x ij is a connection state variable, which takes a value of 0 or 1, indicating whether data centers i and j are connected; β is a bandwidth utilization weight parameter, which is used to balance the requirements of minimizing latency and maximizing bandwidth.

[0157] To ensure that the adjusted network topology meets the actual needs of the data center, the following constraints need to be added to the objective function:

[0158] 1. Link connection constraints:

[0159]

[0160] This means that each data center needs to maintain network connectivity with at least one other data center.

[0161] 2. Bandwidth limitation constraints:

[0162]

[0163] Among them B ij To maximize bandwidth capacity between data centers, ensure that network adjustments do not exceed bandwidth limitations.

[0164] 3. Task transmission delay constraints:

[0165]

[0166] where t k represents the total transmission delay of the kth task, T max is the maximum tolerable delay of the task.

[0167] Preferably, in an embodiment of the present invention, when the proportion of real-time tasks in the network is high, the bandwidth weight β can be set to 0.8, giving priority to low-latency network connections; if the task is more dependent on bandwidth, β can be increased to 0.9. Through the above optimization strategy, network latency can be significantly reduced, while maximizing bandwidth utilization, providing more efficient network support for data center task scheduling.

[0168] In a preferred embodiment of the present invention, an exception handling step is further included. When the resource utilization of any data center exceeds a preset threshold, the task migration path is dynamically adjusted and the resource allocation matrix is ​​recalculated.

[0169] In a preferred embodiment of the present invention, to ensure the stability and efficiency of the system, the exception handling step monitors the resource utilization of the data center in real time. When the resource utilization of a data center exceeds a preset threshold, the dynamic adjustment of the task migration path and the recalculation of the resource allocation matrix are triggered.

[0170] Preferably, the resource utilization threshold U threshold Dynamically set based on task type and data center load. For example:

[0171] For tasks with high real-time requirements, the CPU and GPU utilization thresholds can be set to 80%;

[0172] For common tasks, the utilization threshold of storage resources can be increased to 90%.

[0173] Assume that the current resource usage of data center A is as follows:

[0174] CPU utilization: 85%

[0175] Storage utilization: 70%.

[0176] When the CPU utilization exceeds the threshold of 80%, the exception handling step is triggered.

[0177] After an exception is triggered, the task migration path is dynamically adjusted to reduce the load pressure of the data center. The objective function of recalculating the task migration path matrix is:

[0178]

[0179] Where γ is the load balancing coefficient u j is the current load utilization of the target data center j; by introducing γ·u j, ensure that tasks are migrated to low-load data centers and improve the overall balance of the system.

[0180] After adjusting the task migration path, the optimized resource allocation matrix R′ is recalculated. The optimization objective function J described in this step is iteratively updated to ensure that the new resource allocation scheme meets the global optimal requirements.

[0181] The present invention also discloses an international data center resource scheduling system based on deep reinforcement learning, including:

[0182] Data collection module 1, used to collect the resource status matrix, task requirement matrix and energy consumption matrix of the international data center;

[0183] Data processing module 2, used to construct resource scheduling benefit function and optimize resource allocation strategy based on the collected data;

[0184] Scheduling module 3, used to determine the task migration path based on the optimized resource allocation strategy;

[0185] The output module 4 is used to output the optimized resource allocation matrix and task migration path.

[0186] In a preferred embodiment of the present invention, the international data center resource scheduling system is composed of the following main modules:

[0187] Data acquisition module 1 is responsible for collecting multi-dimensional data related to task scheduling, including:

[0188] Resource status matrix R, describing the resource usage of the data center;

[0189] The task demand matrix D represents the specific requirements of the task for each resource;

[0190] The energy consumption matrix E provides the unit energy consumption information of each data center resource.

[0191] This module obtains the latest data through real-time monitoring and regular synchronization to ensure the accuracy of subsequent processing.

[0192] In the embodiment of the present invention, the data processing module 2 constructs a resource scheduling benefit function based on the collected data, and provides an optimal resource allocation strategy by optimizing the objective function J. The module includes:

[0193] The benefit function construction submodule generates the resource scheduling benefit function F, which is used to quantify the pros and cons of the current resource allocation strategy;

[0194] The optimization submodule generates an optimized resource allocation matrix R′ through iterative optimization.

[0195] The scheduling module 3 determines the migration path of the task based on the optimized resource allocation matrix. Preferably, the module adopts a dynamic adjustment strategy, comprehensively considering network delay, energy consumption cost and load balance.

[0196] The output module 4 provides the optimization results in the form of a matrix, including: the optimized resource allocation matrix R′ and the task migration path matrix P.

[0197] Through the collaborative work of the above modules, this system can efficiently complete the resource scheduling of the international data center and provide reliable support for practical applications.

[0198] The method and system of the present invention fully demonstrate the resource scheduling optimization process based on deep reinforcement learning. The technical solution is efficient, robust and scalable, and provides an innovative solution for the task management of the international data center.

[0199] The international data center resource scheduling method and system based on deep reinforcement learning of the present invention designs a set of optimization strategies for the actual needs of dynamic resource allocation and task migration. In order to verify the superiority of the present invention, the following indicators were selected for testing:

[0200] 1. Resource Utilization (RU): indicates the average utilization of data center resources, expressed in percentage (%).

[0201] 2. Task Completion Time (TCT): It indicates the average time taken for all tasks to be completed from the start to the completion, in seconds (s).

[0202] 3. System Energy Consumption (SEC): It refers to the total energy consumed during the operation of the data center, in kilowatt-hours (kWh).

[0203] 4. Task Migration Cost (TMC): It indicates the average cost of migrating tasks between data centers, in cost units (CU).

[0204] The testing standards and methods are as follows:

[0205] 1. Resource utilization: By monitoring the resource utilization of each data center, the average value is calculated. The detection standard is that a rate higher than 85% indicates that the resources are fully utilized.

[0206] 2. Task completion time: record the time from the start of task assignment to completion, with the best detection standard being less than 20 seconds.

[0207] 3. System energy consumption: Record the operating energy consumption of all data centers. The detection standard is 90% lower than the control ratio, which is considered significant energy saving.

[0208] 4. Task migration cost: record the energy consumption and delay cost of task migration. The detection standard is 80% lower than the control ratio, which is excellent.

[0209] Embodiment: Using the method and system of the present invention, based on deep reinforcement learning resource allocation and task migration strategy, the objective function is optimized and the network topology is dynamically adjusted.

[0210] Comparative example: Using the traditional static rule resource allocation algorithm, a fixed task allocation method is adopted without dynamic adjustment.

[0211] The test task volume is 1,000, including 70% real-time tasks and 30% ordinary tasks. The number of data centers is 5. The resource types of each data center include CPU, GPU and storage. The network delay fluctuates according to the actual environment.

[0212] The test results are summarized in the following table:

[0213] index Example Comparative Example Advantages Resource utilization (RU,%) 92.3 78.5 Improved by 17.6%, more efficient use of resources Task completion time (TCT, s) 15.7 23.4 33% shorter, more efficient task processing System energy consumption (SEC, kWh) 185.2 235.8 Reduced by 21.5%, significant energy saving Task migration cost (TMC, CU) 58.3 97.6 Reduced by 40.3%, more efficient migration

[0214] The analysis and explanation are as follows:

[0215] 1. Improvement of resource utilization: In the embodiment, the deep reinforcement learning algorithm can dynamically optimize the resource allocation strategy, make full use of the remaining resources of the data center, and increase the resource utilization to 92.3%. The comparison ratio adopts fixed rule allocation, which cannot respond to load changes in real time, and has serious resource waste, which is only 78.5%. This result proves that the method of the present invention significantly improves the utilization efficiency of resources, especially in scenarios with complex task distribution.

[0216] 2. Shortening of task completion time: The method of the present invention significantly reduces task waiting time and processing delay by optimizing task migration paths and dynamically adjusting resource allocation, and the average task completion time is shortened to 15.7 seconds. In contrast, due to the lack of dynamic scheduling, the average task completion time is as high as 23.4 seconds. This shows that the method of the present invention can greatly improve task processing efficiency in high real-time task scenarios.

[0217] 3. Reduction of system energy consumption: By comprehensively optimizing the resource allocation matrix and network topology, the method of the present invention significantly reduces the overall energy consumption of the system operation to only 185.2 kWh. The comparison ratio has a high energy consumption of 235.8 kWh due to the fixed allocation strategy. The energy-saving effect fully demonstrates the sustainability and economy of the method of the present invention in large-scale data center scenarios.

[0218] 4. Reduction of task migration cost: In the embodiment, the task migration cost is reduced to 58.3CU, which is much lower than 97.6CU in the comparison. This is mainly due to the migration path and bandwidth utilization strategy optimized by deep reinforcement learning, which reduces the migration delay and energy consumption cost. The method of the present invention can efficiently process dynamic tasks, and the migration is more accurate and economical.

[0219] Based on the above results, the method of the present invention shows significant advantages in all test indicators, especially in terms of resource utilization and task completion time, which fully reflects the synergistic effect of dynamic optimization and deep reinforcement learning. Through dynamic resource allocation, migration path optimization and network topology adjustment, the method of the present invention successfully solves dynamic environment problems that traditional methods cannot cope with, and provides an efficient and intelligent solution for international data center resource scheduling.

[0220] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An international data center resource scheduling method based on deep reinforcement learning, characterized in that: include: The acquisition steps include: Obtain a resource status matrix of the international data center, where the resource status matrix includes resource utilization, computing capacity, and network load status of each data center; Obtain a task requirement matrix, where the task requirement matrix includes the computing requirement, storage requirement, and latency requirement of the task; Obtain an energy consumption matrix, where the energy consumption matrix includes unit resource energy consumption of each data center; Processing steps include: Constructing a resource scheduling benefit function based on the resource status matrix and the task requirement matrix; Optimizing the resource allocation strategy according to the energy consumption matrix and the resource scheduling benefit function, and generating an optimized resource allocation matrix; Based on the optimized resource allocation matrix, determine the task migration path and generate a task migration matrix; Output steps include: The optimized resource allocation matrix and the task migration matrix are outputted to realize resource optimization and task scheduling of the international data center.

2. The method according to claim 1, characterized in that The obtaining step further comprises: Obtain a network delay matrix, where the network delay matrix includes network transmission delays between data centers; The network delay matrix is ​​used as an input for optimizing resource allocation strategy to comprehensively consider the network delay cost of task migration.

3. The method according to claim 1, characterized in that The processing steps specifically include: Based on the resource status matrix and task requirement matrix, a resource scheduling benefit function F is constructed: Where: F is the resource scheduling benefit value; w ij is a weight parameter, indicating the importance of task j in data center i; r ij is the resource utilization rate, which indicates the resource consumption ratio of task j in data center i.

4. The method according to claim 3, characterized in that The optimized resource allocation strategy specifically includes: Minimize resource energy consumption by optimizing the objective function J: Where: E i is the unit resource energy consumption of the ith data center, C i is the total computing capacity of the ith data center; λ is the regularization coefficient, which controls the scheduling variation; is the Frobenius norm, which is used to limit the difference between the optimized resource matrix and the original matrix.

5. The method according to claim 4, characterized in that The optimization of the task migration path is based on the following objective function: Where: P is the task migration path matrix; p ij is the number of migration tasks from data center i to j; c ij The cost of task migration includes network latency and energy consumption.

6. The method according to claim 5, characterized in that The task migration path further satisfies the following constraints: It indicates that the inflow and outflow of task migration are balanced.

7. The method according to claim 1, characterized in that The optimized resource allocation matrix is ​​further combined with the network topology adjustment strategy, using the following comprehensive optimization objective function: Among them, Z is the comprehensive optimization benefit, and μ is the weight coefficient of task migration cost.

8. The method according to claim 7, characterized in that The network topology adjustment strategy is optimized based on an integer linear programming problem, wherein the optimization objectives are to minimize network delay and maximize bandwidth utilization.

9. The method according to claim 8, characterized in that It further includes an exception handling step, when the resource utilization of any data center exceeds a preset threshold, dynamically adjusts the task migration path and recalculates the resource allocation matrix.

10. An international data center resource scheduling system based on deep reinforcement learning that executes the method according to any one of claims 1 to 9, characterized in that: include: Data collection module, used to collect the resource status matrix, task demand matrix and energy consumption matrix of the international data center; Data processing module, used to construct resource scheduling benefit function and optimize resource allocation strategy based on collected data; The scheduling module is used to determine the task migration path based on the optimized resource allocation strategy; The output module is used to output the optimized resource allocation matrix and task migration path.

Images