Server computing power resource dynamic scheduling method based on multi-objective optimization

Abstract

The invention discloses a server computing power resource dynamic scheduling method based on multi-objective optimization, and relates to the technical field of server scheduling, and the method comprises the steps: obtaining real-time operation state data of a server cluster and attribute information of a to-be-scheduled computing task in response to a dynamic scheduling trigger event; based on the real-time operation state data and the attribute information, constructing a multi-objective optimization model; the multi-objective optimization model is solved, and a Pareto optimal solution set representing different resource allocation schemes is obtained; and selecting an optimal scheduling scheme from the Pareto optimal solution set based on a preset operation strategy. The method has the advantages that the dynamic multi-objective optimization model fusing the performance, the energy efficiency and the resource utilization rate is constructed, and online solving is performed by adopting the decomposition-based evolutionary algorithm, so that comprehensive optimization and optimal balance of the computing power resource utilization rate, the task processing efficiency and the overall energy efficiency of the system are realized on the premise of ensuring the service level agreement.

Description

Technical Field

[0001] This invention relates to the field of server scheduling technology, specifically to a dynamic scheduling method for server computing resources based on multi-objective optimization. Background Technology

[0002] With the rapid expansion of cloud computing, edge computing, and data center scale, the scheduling of computing resources in server clusters faces unprecedented complexity challenges. Most existing scheduling methods optimize for a single performance metric (such as shortest response time or highest resource utilization), making it difficult to simultaneously address multiple objectives such as energy efficiency, cost, and service quality. In environments with highly heterogeneous resources and dynamically changing workloads, such single-objective optimization strategies can easily lead to severe resource fragmentation, high energy consumption, or violations of service level agreements, failing to meet the comprehensive needs of modern green data centers and diverse business scenarios.

[0003] Currently, while some research attempts to incorporate multi-objective optimization theory, these methods often employ fixed-weight linear weighting to transform multiple objectives into a single objective, or use computationally complex offline optimization algorithms, lacking the ability to respond to and adapt to dynamic environments in real time. Furthermore, existing solutions often neglect the closed-loop learning between scheduling strategies and actual operational feedback, making it difficult for the model to continuously adapt to changes in load patterns and limiting its long-term effectiveness and robustness in real-world production systems. Therefore, there is an urgent need for an intelligent scheduling method for server computing resources that can respond online in real time, dynamically balance multiple conflicting objectives, and possess self-learning capabilities. Summary of the Invention

[0004] To address the aforementioned technical problems, a dynamic scheduling method for server computing resources based on multi-objective optimization is provided. This technical solution solves at least one of the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for dynamic scheduling of server computing resources based on multi-objective optimization includes:

[0007] In response to dynamic scheduling trigger events, obtain real-time running status data of the server cluster and attribute information of the computing tasks to be scheduled;

[0008] Based on the real-time running status data and the attribute information, a multi-objective optimization model is constructed with the optimization objectives of minimizing task completion time, minimizing total cluster energy consumption, and minimizing resource fragmentation rate, and with resource capacity and task allocation as constraints.

[0009] A decomposition-based multi-objective evolutionary algorithm is used to solve the multi-objective optimization model to obtain a Pareto optimal solution set representing different resource allocation schemes;

[0010] Based on a preset operation strategy, an optimal scheduling scheme is selected from the Pareto optimal solution set, and the computing tasks to be scheduled are dynamically allocated to the corresponding server nodes in the server cluster for execution according to the scheme.

[0011] Preferably, the response to the dynamic scheduling trigger event specifically includes:

[0012] The dynamic scheduling trigger events include new computing task arrival events, server node status abnormality events, and periodic rescheduling events.

[0013] In the event of the arrival of a new computing task, the computing resource requirements, storage resource requirements, network bandwidth requirements, and task priority or deadline of the newly arrived computing task are obtained.

[0014] During the periodic rescheduling event, the attribute information of all computing tasks currently waiting to be executed in the queue is obtained simultaneously to perform global resource reallocation optimization.

[0015] Preferably, the construction of the multi-objective optimization model specifically includes:

[0016] The objective function for minimizing task completion time in the optimization objective is defined by calculating the sum of the times from arrival to completion of all tasks;

[0017] The objective function for minimizing the total energy consumption of the cluster is defined by integrating the real-time power consumption models of each server node and calculating the total power consumption integral within the scheduling cycle.

[0018] The objective function for minimizing resource fragmentation rate is defined by calculating the sum of squares or coefficient of variation of unused resources on all server nodes, in order to simultaneously promote load balancing and resource consolidation.

[0019] Preferably, the step of using a decomposition-based multi-objective evolutionary algorithm to solve the multi-objective optimization model and obtain a Pareto optimal solution set representing different resource allocation schemes specifically includes:

[0020] The multi-objective optimization problem is decomposed into a series of single-objective sub-problems using a weight vector, and these sub-problems are optimized simultaneously using an evolutionary algorithm.

[0021] In each generation of evolution, optimization information is exchanged between adjacent subproblems through domain relationships, and Pareto fronts are searched collaboratively.

[0022] An incremental optimization mechanism is adopted, in which the optimized solutions of historical scheduling cycles are used as part of the initial population to accelerate the convergence speed of the current cycle and improve the quality of the solutions.

[0023] Preferably, the selection of an optimal scheduling scheme from the Pareto optimal solution set based on a preset operating strategy specifically includes:

[0024] The operational strategy is dynamically set by the system administrator according to actual business needs, including performance-first mode, energy-efficiency-first mode, or cost-first mode.

[0025] In the selected mode, the approximation ideal solution ranking method is used to comprehensively evaluate and rank the various schemes in the Pareto optimal solution set;

[0026] The scheduling scheme with the highest comprehensive evaluation score is selected as the final execution scheme and sent to the resource management platform of the server cluster.

[0027] Preferably, the method further includes a feedback learning process parallel to the dynamic scheduling process, the feedback learning process including:

[0028] After each scheduling execution, the actual task execution time, resource utilization, and energy consumption data are collected as feedback information.

[0029] The feedback information is compared with the prediction information before scheduling to calibrate the prediction model parameters of the load prediction module.

[0030] Based on long-term feedback data, the weight coefficients of each objective function in the multi-objective optimization model are adaptively adjusted using reinforcement learning methods, so that the scheduling strategy can continuously adapt to changes in load patterns.

[0031] Preferably, the method further includes preprocessing the computational task, specifically including:

[0032] Based on the characteristics of computing tasks' consumption of computing, storage, network, and memory resources, they are classified into several categories, including computing-intensive, data-intensive, and memory-intensive.

[0033] Based on the hardware configuration and performance characteristics of the server nodes, they are marked as preferred nodes suitable for handling specific task categories;

[0034] When constructing the multi-objective optimization model, a positive incentive coefficient is introduced into the decision variable that assigns tasks to their preferred nodes in order to guide the optimization process to achieve feature matching between tasks and resources.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention constructs a dynamic multi-objective optimization model that integrates performance, energy efficiency, and resource utilization, and employs a decomposition-based evolutionary algorithm for online solution. This model provides a series of Pareto-optimal scheduling schemes for dynamically changing server clusters and workloads, thereby achieving comprehensive optimization and optimal trade-offs between computing resource utilization, task processing efficiency, and overall system energy efficiency while ensuring service level agreements (SLAs). Furthermore, this method integrates feedback learning and adaptive adjustment mechanisms, enabling the scheduling strategy to continuously track and adapt to changes in load patterns and system states. This significantly improves the intelligence level, environmental adaptability, and long-term operational economy of resource management in large-scale heterogeneous data centers. Attached Figure Description

[0037] Figure 1 The flowchart shows the dynamic scheduling method for server computing resources based on multi-objective optimization proposed in this scheme.

[0038] Figure 2 This is an architecture diagram of the electronic devices in this solution;

[0039] Figure 3 This is a schematic diagram of the computer-readable storage medium structure in this scheme.

[0040] The numbers on the map are:

[0041] 500 - Electronic device; 501 - Bus; 502 - CPU; 503 - ROM; 504 - RAM; 505 - Communication port; 506 - Input / output component; 507 - Hard disk; 508 - User interface; 600 - Computer-readable storage medium. Detailed Implementation

[0042] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0043] Reference Figure 1 As shown, a method for dynamic scheduling of server computing resources based on multi-objective optimization includes:

[0044] In response to dynamic scheduling trigger events, the system acquires real-time operational status data of the server cluster and attribute information of the computational tasks to be scheduled. Dynamic scheduling trigger events aim to ensure that the method can respond promptly to changes within and outside the system, rather than executing at fixed intervals, thereby improving scheduling agility. The acquired real-time operational status data, such as CPU / memory utilization, real-time power consumption, and network bandwidth usage of each node, along with task attribute information such as resource requirements, priorities, and dependencies, together form a snapshot depicting the current instantaneous state of the system and the load to be processed, providing accurate and comprehensive input for subsequent optimization modeling.

[0045] Based on real-time operational status data and attribute information, a multi-objective optimization model is constructed with the objectives of minimizing task completion time, minimizing total cluster energy consumption, and minimizing resource fragmentation rate, constrained by resource capacity and task allocation. By simultaneously setting three competing objectives—task completion time, total cluster energy consumption, and resource fragmentation rate—the model can systematically and quantitatively characterize the trade-offs between performance, energy efficiency, and resource utilization. Constraints (such as single-task single-server and resource capacity limits) ensure that any generated scheduling scheme must be physically executable, thus standardizing the complex scheduling problem into a structured multi-objective constrained optimization problem.

[0046] A decomposition-based multi-objective evolutionary algorithm is employed to solve a multi-objective optimization model, obtaining a Pareto optimal solution set representing different resource allocation schemes. This algorithm efficiently approximates the entire Pareto front by decomposing the original problem into a set of related single-objective subproblems and co-evolving them. Its advantage lies in obtaining a set of Pareto optimal solutions in a single run, rather than a single solution. Each solution in this set represents a different trade-off preference among multiple objectives (e.g., one solution favors low latency, while another favors high energy efficiency), thus providing a rich set of non-dominated candidate solutions for the final decision.

[0047] Based on a pre-defined operational strategy, an optimal scheduling scheme is selected from the Pareto optimal solution set. The computational tasks to be scheduled are then dynamically allocated to the corresponding server nodes in the server cluster according to this scheme. The pre-defined operational strategy, such as prioritizing performance during peak hours or prioritizing energy consumption at night, reflects the administrator's business intentions, while the Pareto solution set provides various technical possibilities for achieving these intentions. A decision criterion, such as the TOPSIS method, is used to select the scheme that best fits the current strategy from the solution set, thus combining business objectives with technical optimization. The final allocation instruction is sent to the cluster management system to complete the actual allocation of computing resources, forming a complete decision-making-execution closed loop.

[0048] Specifically, the responses to dynamically scheduled trigger events in this solution include:

[0049] Dynamic scheduling trigger events include new computing task arrival events, server node status abnormality events, and periodic rescheduling events;

[0050] In the event of a new computing task arriving, obtain the computing resource requirements, storage resource requirements, network bandwidth requirements, and task priority or deadline of the newly arriving computing task.

[0051] During periodic rescheduling events, the attribute information of all computing tasks currently waiting to be executed in the queue is obtained simultaneously to perform global resource reallocation optimization.

[0052] In practical implementation, the dynamic scheduling module uses both event listening and periodic timing mechanisms to capture and respond to different scheduling trigger events in real time. When a new computing task arrives, the system listens to the task queue and immediately parses its submission description file to obtain its computing resource requirements, storage resource requirements, network bandwidth requirements, and task priority or deadline, which serve as the core input for the task to be scheduled. When the monitoring system detects abnormal server node status, such as node failure, resource overload, or network partitioning, this event will immediately trigger the rescheduling and computation of the affected tasks. In addition, the system has a built-in independent rescheduling timer that starts a global optimization process at a fixed period (e.g., every 5 minutes) regardless of external events. During this periodic event, the system pulls all tasks currently waiting in the queue and their attributes, and combines this with the latest cluster status to perform a round of global resource reassessment and allocation optimization to eliminate resource fragmentation or uneven load that may be caused by incremental scheduling. This hybrid triggering mechanism ensures that the system can respond quickly to immediate changes while continuously maintaining the long-term health and efficiency of the resource pool through periodic global optimization.

[0053] In this scheme, the construction of a multi-objective optimization model specifically includes:

[0054] The optimization objective function, which minimizes the task completion time, is defined by calculating the sum of the times from arrival to completion for all tasks. The mathematical expression for this time objective function is:

[0055] ;

[0056] The time objective function value, Let be the arrival time of the i-th task. Let i be the completion time of the i-th task.

[0057] The objective function for minimizing the total energy consumption of the cluster is defined by integrating the real-time power consumption models of each server node and calculating the integral of the total power consumption within the scheduling cycle. The mathematical expression for the objective function for the total energy consumption of the cluster is as follows:

[0058] ;

[0059] The objective function value for the total energy consumption of the cluster. The power consumption of server j at time t

[0060] The objective function for minimizing resource fragmentation rate is defined by calculating the sum of squares or coefficient of variation of unused resources on all server nodes, aiming to simultaneously promote load balancing and resource consolidation. The mathematical expression for the objective function for resource fragmentation rate is:

[0061] ;

[0062] The objective function value is the resource fragmentation rate. For the resources already used by server j, This represents the total resources of server j.

[0063] Constructing a multi-objective optimization model is the core mathematical foundation for achieving intelligent scheduling. This process is automatically completed by the modeling module in the scheduling engine. The module first precisely quantifies three core optimization objectives based on real-time data obtained from the monitoring system:

[0064] Time objective function: The module reads the system record arrival time for each task i to be scheduled. Based on the assumed resource allocation scheme for task i (such as allocation to a specific server j), and using the server's current load, historical execution time models of task i's own computational requirements, or benchmark data, the completion time is estimated. . It is the sum of the differences between the estimated completion time and the arrival time of all related tasks, and its value directly reflects the impact of the scheduling scheme on the overall task processing delay.

[0065] Energy consumption objective function: The module calls a pre-built server power consumption model library. For each server node j, the power consumption model... It is a multivariate function of CPU utilization, memory usage, disk I / O, and operating frequency. When evaluating a scheduling scheme, the module dynamically predicts the change in the power consumption curve of each server j within a future scheduling cycle T based on the additional load added to each server j by the scheme, and calculates the total energy consumption within that cycle through numerical integration. It is the sum of the power consumption integrals of all servers in the cluster within period T, used to measure the energy efficiency level of the solution.

[0066] Resource fragmentation rate objective function: The module abstracts the resources of server j (such as the number of CPU cores and the number of GB of memory) into a multi-dimensional resource vector. The allocated resources are . The computational design ensures that the penalty term for a server's contribution increases quadratically as its resource utilization decreases, thus strongly driving the optimization algorithm to avoid generating low-utilization nodes, effectively integrating resources and improving load balancing.

[0067] After defining the three competing objective functions, the modeling module systematically adds constraints, primarily including: the sum of the computation, memory, and storage requirements of each task must not exceed the corresponding resource capacity of the target server to which it is assigned; and each task must be executed on only one server. Finally, this multi-objective constrained optimization problem is formally submitted to the subsequent solution algorithm module.

[0068] In this scheme, a decomposition-based multi-objective evolutionary algorithm is used to solve the multi-objective optimization model, obtaining a Pareto optimal solution set representing different resource allocation schemes, specifically including:

[0069] The multi-objective optimization problem is decomposed into a series of single-objective sub-problems using a weight vector, and these sub-problems are optimized simultaneously using an evolutionary algorithm;

[0070] In each generation of evolution, optimization information is exchanged between adjacent subproblems through domain relationships, and Pareto fronts are searched collaboratively.

[0071] An incremental optimization mechanism is adopted, in which the optimized solutions of historical scheduling cycles are used as part of the initial population to accelerate the convergence speed of the current cycle and improve the quality of the solutions.

[0072] In the specific implementation process, the original multi-objective optimization problem is first decomposed into N single-objective subproblems. This decomposition is achieved by generating a set of uniformly distributed weight vectors, where the sum of all components in each weight vector is 1 and all components are non-negative. Each weight vector corresponds to a scalarized subproblem, and its fitness function is a weighted Chebyshev function. Simultaneously, the algorithm defines a neighborhood for each subproblem, consisting of the T subproblems with the closest Euclidean distance between the weight vectors, to promote subsequent co-evolution. The initial population is generated through two methods: random generation and incorporating historical optimal solutions. The historical optimal solutions come from the Pareto optimal solution set saved in the previous scheduling cycle. This incremental optimization mechanism significantly improves the quality of the initial solutions and the convergence speed of the algorithm.

[0073] In each generation of evolution, the algorithm optimizes each subproblem sequentially. For the current subproblem, two parent individuals are randomly selected from its neighborhood, and a new child individual is generated through simulated binary crossover and polynomial mutation operations. Subsequently, for this child individual, the feasibility of the resource allocation scheme is improved, for example, ensuring the integrity of task allocation and satisfying resource constraints. Next, the objective function value of the child individual is calculated, and the ideal reference point is updated accordingly. Then, the algorithm uses the child individual to update the neighborhood of the current subproblem. This mechanism of sharing information about excellent individuals through neighborhood relationships enables the various subproblems to collaboratively explore the entire target space.

[0074] The evolutionary process is repeated until a preset termination condition is met (such as reaching the maximum number of iterations or solution set convergence). Finally, the algorithm extracts all non-dominated solutions from the current optimal solutions of all subproblems, forming an approximately Pareto optimal solution set. Each solution in this set represents a feasible resource allocation scheme, exhibiting different trade-offs among the three objectives of task completion time, total cluster energy consumption, and resource fragmentation rate. This solution set provides system administrators with a rich set of decision-making options, thus realizing a paradigm shift from a single optimal solution to a series of optimal trade-off schemes, laying a solid foundation for subsequent final decisions based on operational strategies.

[0075] In this scheme, selecting an optimal scheduling scheme from the Pareto optimal solution set based on a preset operating strategy specifically includes:

[0076] The operational strategy is dynamically set by the system administrator based on actual business needs, including performance-first mode, energy-efficiency-first mode, or cost-first mode.

[0077] Under the selected mode, the approximation ideal solution ranking method is used to comprehensively evaluate and rank the various schemes in the Pareto optimal solution set;

[0078] The scheduling scheme with the highest comprehensive evaluation score is selected as the final execution scheme and then sent to the resource management platform of the server cluster.

[0079] In its implementation, this invention employs a dedicated decision-making module to execute the decision-making process from the Pareto optimal solution set to the final scheduling scheme. At the core of this module is a configurable strategy engine, allowing system administrators to dynamically select preset operational strategies such as performance-first, energy-efficiency-first, or cost-first through a management interface. Each strategy has a pre-defined set of target weight vectors and normalization methods within the engine; for example, in performance-first mode, the weight of the task completion time target is significantly increased, while the normalization range of energy consumption data is appropriately compressed to ensure that performance indicators dominate the ranking process.

[0080] After selecting a strategy, the decision-making module uses the approximation-ideal-solution ranking method to automatically evaluate and rank the Pareto solution set. The specific steps are as follows: First, vector normalization is performed on the values ​​of all schemes in the solution set on the three objective functions to eliminate the influence of dimensions. Next, the weighted evaluation value of each scheme is calculated based on the weight of the current strategy. Then, the positive and negative ideal solutions in the three objective directions are determined respectively. Subsequently, the Euclidean distance between each actual scheme and the positive and negative ideal solutions is calculated. Finally, the Pareto distance is determined using the formula... Calculate the overall evaluation score for each option. ,in and These are the distances to the positive and negative ideal solutions, respectively. (Score) The higher the value (the closer to 1), the closer the solution is to the ideal solution.

[0081] The decision-making module automatically selects the scheme with the highest comprehensive evaluation score as the optimal scheduling scheme for the current scheduling cycle. This scheme is then converted into a scheduling instruction or resource binding description that the resource management platform can recognize, and executed through the platform's API interface. This process transforms abstract, multi-objective mathematical optimization results into a clear, executable operation instruction that best aligns with the current business intent, completing the final step from multi-objective trade-offs to business decision-making.

[0082] In some preferred embodiments, the method further includes a feedback learning process parallel to the dynamic scheduling process, the feedback learning process including:

[0083] After each scheduling execution, the actual task execution time, resource utilization, and energy consumption data are collected as feedback information.

[0084] The feedback information is compared with the prediction information before scheduling to calibrate the prediction model parameters of the load prediction module.

[0085] Based on long-term feedback data, reinforcement learning methods are used to adaptively adjust the weight coefficients of each objective function in the multi-objective optimization model, so that the scheduling strategy can continuously adapt to changes in load patterns.

[0086] Specifically, this loop acts as a background service, automatically activating after each scheduling plan is executed. Its core is a data collector, which continuously gathers data on the actual execution performance of the plan from the cluster monitoring system and task logs. This includes: the actual start and end times of each task, the real power consumption curves of each server node within the scheduling cycle (obtained through out-of-band management interfaces such as IPMI), and real-time utilization snapshots of various resources. This actual data, along with the predicted data used in scheduling decisions (such as estimated task completion times and predicted energy consumption), is fed into the feedback analysis module, forming a feedback information pair for learning and calibration.

[0087] The feedback analysis module employs an incremental learning mechanism to process this feedback information. For the load prediction model, the module compares the actual task execution time with the predicted time, and uses the error sequence to fine-tune the weight parameters of the prediction model (such as an LSTM neural network) using an online gradient descent algorithm, making its predictions of similar future tasks more accurate. For system-level long-term adaptation, this embodiment integrates a reinforcement learning method based on policy gradients. This method treats the weight coefficients of each objective function in the multi-objective optimization model as adjustable policies. The feedback analysis module tracks the actual comprehensive benefits corresponding to the scheduling schemes generated under different weight configurations over a long period, for example, a composite index combining performance achievement rate, unit computing energy consumption, and resource utilization. By learning the mapping relationship between historical decisions and long-term benefits, the reinforcement learning agent periodically outputs weight adjustment suggestions, aiming to maximize long-term benefits. These adjusted weights are dynamically injected into the multi-objective optimization model in subsequent scheduling cycles, enabling the entire scheduling strategy to automatically adapt to seasonal changes in business load, hardware performance degradation, or gradual changes in operational strategies, achieving a leap from static optimization to dynamic self-evolution.

[0088] In some preferred embodiments, the method further includes preprocessing the computational task, specifically including:

[0089] Based on the characteristics of computing tasks' consumption of computing, storage, network, and memory resources, they are classified into several categories, including computing-intensive, data-intensive, and memory-intensive.

[0090] Based on the hardware configuration and performance characteristics of the server nodes, they are marked as preferred nodes suitable for handling specific task categories;

[0091] When constructing a multi-objective optimization model, positive incentive coefficients are introduced into the decision variables that assign tasks to their preferred nodes to guide the optimization process and achieve feature matching between tasks and resources.

[0092] Specifically, before building the optimization model, an intelligent preprocessing stage is introduced to achieve a fine-grained match between task requirements and server characteristics, thereby improving overall scheduling efficiency. This stage is completed collaboratively by a task feature analyzer and a node profile builder.

[0093] The task feature analyzer analyzes and classifies the computational tasks entering the system. The analyzer extracts resource consumption characteristics based on the resource request specifications declared when the task is submitted, performance profiles of similar historical tasks, or through lightweight dynamic analysis, such as sampling the behavior of the task in its initial stage. Based on these characteristics, tasks are categorized into predefined classes, such as: computationally intensive tasks that consume large amounts of CPU or GPU computing power, such as scientific computing and model training; data-intensive tasks that frequently involve large-scale disk I / O or network transfers, such as big data processing and video transcoding; and memory-intensive tasks that require large amounts of memory space or high memory bandwidth, such as in-memory databases and real-time analytics.

[0094] Meanwhile, the node profiling builder creates a feature profile for each server node in the cluster. This profile is based on the node's hardware configuration, such as CPU model and number of cores, GPU model and number, memory capacity and channels, storage type and IOPS, network bandwidth, and benchmark performance test data. Based on this profile, nodes are dynamically labeled as preferred nodes for handling specific types of tasks. For example, nodes equipped with high-performance GPUs are labeled as compute-intensive preferred nodes; nodes equipped with high-speed NVMe SSD arrays and high-speed networks are labeled as data-intensive preferred nodes; and nodes equipped with large amounts of memory are labeled as memory-intensive preferred nodes.

[0095] When constructing the multi-objective optimization model, the system encodes this preprocessed information into the objective function. Specifically, when defining the evaluation criteria for task allocation schemes, the model introduces a positive incentive term for the decision to allocate tasks to its preferred nodes. For example, a reward term is added to the objective function: ,in It is a decision variable. It is a coefficient that determines when the type of task i matches the preference label of node j. Take a large positive value, otherwise 0 or a small positive value; α is a positive coefficient that adjusts the intensity of the incentive. This incentive term makes the optimization algorithm tend to generate solutions that can place the right tasks on the right nodes when searching for the Pareto front, thereby effectively reducing resource contention, reducing task execution time, and indirectly improving energy efficiency, achieving better resource orchestration based on semantic understanding.

[0096] Furthermore, the method according to the embodiments of this application can also be achieved by means of... Figure 2 The architecture of the electronic device shown is used to implement this. For example... Figure 2As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, ROM 503, RAM 504, a communication port 505 connected to a network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as ROM 503 or hard disk 507, may store a server computing resource dynamic scheduling method based on multi-objective optimization provided in this application. The electronic device 500 may also include a user interface 508. Of course, Figure 2 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 2 One or more components in the illustrated electronic device.

[0097] Figure 3 This is a schematic diagram of a computer-readable storage medium structure provided in one embodiment of this application. Figure 3 The diagram illustrates a computer-readable storage medium 600 according to one embodiment of this application. The computer-readable storage medium 600 stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a server computing resource dynamic scheduling method based on multi-objective optimization according to an embodiment of this application, as described with reference to the above figures. The computer-readable storage medium 600 includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0098] In summary, the advantages of this invention are as follows: by constructing a dynamic multi-objective optimization model that integrates performance, energy efficiency, and resource utilization, and by using a decomposition-based evolutionary algorithm for online solution, it can provide a series of Pareto optimal scheduling schemes for dynamically changing server clusters and workloads, thereby achieving comprehensive optimization and optimal balance between computing resource utilization, task processing efficiency, and overall system energy efficiency while ensuring service level agreements.

[0099] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for dynamic scheduling of server computing resources based on multi-objective optimization, characterized in that, include: In response to dynamic scheduling trigger events, obtain real-time running status data of the server cluster and attribute information of the computing tasks to be scheduled; Based on the real-time running status data and the attribute information, a multi-objective optimization model is constructed with the optimization objectives of minimizing task completion time, minimizing total cluster energy consumption, and minimizing resource fragmentation rate, and with resource capacity and task allocation as constraints. A decomposition-based multi-objective evolutionary algorithm is used to solve the multi-objective optimization model to obtain a Pareto optimal solution set representing different resource allocation schemes; Based on a preset operation strategy, an optimal scheduling scheme is selected from the Pareto optimal solution set, and the computing tasks to be scheduled are dynamically allocated to the corresponding server nodes in the server cluster for execution according to the scheme.

2. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 1, characterized in that, The response to the dynamically scheduled triggering event specifically includes: The dynamic scheduling trigger events include new computing task arrival events, server node status abnormality events, and periodic rescheduling events. In the event of the arrival of a new computing task, the computing resource requirements, storage resource requirements, network bandwidth requirements, and task priority or deadline of the newly arrived computing task are obtained. During the periodic rescheduling event, the attribute information of all computing tasks currently waiting to be executed in the queue is obtained simultaneously to perform global resource reallocation optimization.

3. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 2, characterized in that, The construction of the multi-objective optimization model specifically includes: The objective function for minimizing task completion time in the optimization objective is defined by calculating the sum of the times from arrival to completion of all tasks; The objective function for minimizing the total energy consumption of the cluster is defined by integrating the real-time power consumption models of each server node and calculating the total power consumption integral within the scheduling cycle. The objective function for minimizing resource fragmentation rate is defined by calculating the sum of squares or coefficient of variation of unused resources on all server nodes, in order to simultaneously promote load balancing and resource consolidation.

4. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 3, characterized in that, The step of using a decomposition-based multi-objective evolutionary algorithm to solve the multi-objective optimization model and obtain a Pareto optimal solution set representing different resource allocation schemes specifically includes: The multi-objective optimization problem is decomposed into a series of single-objective sub-problems using a weight vector, and these sub-problems are optimized simultaneously using an evolutionary algorithm. In each generation of evolution, optimization information is exchanged between adjacent subproblems through domain relationships, and Pareto fronts are searched collaboratively. An incremental optimization mechanism is adopted, in which the optimized solutions of historical scheduling cycles are used as part of the initial population to accelerate the convergence speed of the current cycle and improve the quality of the solutions.

5. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 4, characterized in that, The selection of an optimal scheduling scheme from the Pareto optimal solution set based on a preset operating strategy specifically includes: The operational strategy is dynamically set by the system administrator according to actual business needs, including performance-first mode, energy-efficiency-first mode, or cost-first mode. In the selected mode, the approximation ideal solution ranking method is used to comprehensively evaluate and rank the various schemes in the Pareto optimal solution set; The scheduling scheme with the highest comprehensive evaluation score is selected as the final execution scheme and sent to the resource management platform of the server cluster.

6. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 5, characterized in that, It also includes a feedback learning process that runs parallel to the dynamic scheduling process, the feedback learning process comprising: After each scheduling execution, the actual task execution time, resource utilization, and energy consumption data are collected as feedback information. The feedback information is compared with the prediction information before scheduling to calibrate the prediction model parameters of the load prediction module. Based on long-term feedback data, the weight coefficients of each objective function in the multi-objective optimization model are adaptively adjusted using reinforcement learning methods, so that the scheduling strategy can continuously adapt to changes in load patterns.

7. The method for dynamic scheduling of server computing resources based on multi-objective optimization according to claim 6, characterized in that, It also includes preprocessing of the computational task, specifically including: Based on the characteristics of computing tasks' consumption of computing, storage, network, and memory resources, they are classified into several categories, including computing-intensive, data-intensive, and memory-intensive. Based on the hardware configuration and performance characteristics of the server nodes, they are marked as preferred nodes suitable for handling specific task categories; When constructing the multi-objective optimization model, a positive incentive coefficient is introduced into the decision variable that assigns tasks to their preferred nodes in order to guide the optimization process to achieve feature matching between tasks and resources.

8. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a dynamic scheduling method for server computing resources based on multi-objective optimization as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the server computing resource dynamic scheduling method based on multi-objective optimization as described in any one of claims 1-7.

Images