Search

CN-122019322-A - Adaptive hybrid parameter optimization method and device integrating multitask migration and regional potential modeling

CN122019322ACN 122019322 ACN122019322 ACN 122019322ACN-122019322-A

Abstract

The invention discloses a self-adaptive mixed parameter optimization method and a device for fusing multitask migration and regional potential modeling, the method is based on performance feedback obtained in the running process of the system, unified modeling is carried out on the mixed parameter space, and internal differences of the parameter space are described through regional management. By evaluating the historical performance and potential optimality of different parameter regions, the parameter search is directed to adaptively focus on high potential regions, thereby improving search efficiency and reducing ineffective exploration. Meanwhile, a multitask migration mechanism based on parameter searching behavior is introduced, and an initial searching process of a new task is guided by utilizing region evolution and searching characteristics in a historical task, so that optimization convergence is accelerated. The method does not need to interrupt system operation or manual intervention, can effectively utilize cross-task optimization experience, and has good universality and engineering application value.

Inventors

  • LIN WEIWEI
  • LU YIKANG

Assignees

  • 华南理工大学

Dates

Publication Date
20260512
Application Date
20260128

Claims (10)

  1. 1. The adaptive mixed parameter optimization method integrating multitasking migration and regional potential modeling is characterized by comprising the following steps of: The performance monitoring and searching track construction step comprises the steps of collecting performance feedback information corresponding to the current parameter configuration in the running process of a target program or a system, carrying out association storage on the performance feedback information and the corresponding parameter configuration combination, and constructing a parameter-performance historical data set and a parameter searching track for subsequent region modeling and multitasking migration analysis; The mixed parameter space regional modeling and potential evaluation step comprises the steps of carrying out unified representation modeling on a mixed parameter space containing continuous parameters and discrete parameters, dividing the parameter space into a plurality of parameter configuration areas, constructing an area-level proxy model based on historical performance expression and model uncertainty in each parameter configuration area, and calculating corresponding area optimization potential indexes; The self-adaptive area scheduling step based on area potential, namely, according to the optimizing potential evaluation result of each parameter configuration area, self-adaptively executing area holding, expanding, contracting or switching operation, and dynamically determining a target parameter configuration area of the current optimizing stage so as to realize the balance between exploration and utilization; the multi-task migration guiding step based on the searching behavior comprises the steps of analyzing parameter searching behavior, region selection track and performance feedback distribution in a historical task, evaluating the similarity of the historical task and a current task in the searching behavior level, and migrating high-potential region information or region evolution mode of the similar historical task to the current task when the similar historical task exists, wherein the high-potential region information or region evolution mode is used for guiding region selection and sampling decision; And updating the regional proxy model, the regional potential evaluation result and the search behavior characterization based on the performance feedback information, and returning to execute the performance monitoring and search track construction step to form a continuous online self-adaptive optimization process.
  2. 2. The adaptive hybrid parameter optimization method for fusion of multitasking migration and regional potential modeling of claim 1, wherein in the performance monitoring and search trajectory construction step: Collecting performance feedback information by taking a preset time interval or the completion of each parameter execution as a trigger condition; the performance feedback information is expressed as a performance feedback vector , wherein, Represent the first Configuring corresponding performance evaluation results by secondary parameters; And performing associated storage on the performance feedback vector and the corresponding parameter configuration combination to construct a global performance data set for subsequent optimization decision, wherein the global performance data set comprises a parameter-performance history data set and a parameter search track.
  3. 3. The adaptive hybrid parameter optimization method for fusing multitasking migration and regional potential modeling of claim 1, wherein in the regional potential-based adaptive regional scheduling step: determining the optimization potential of the region by calculating the performance expected value and the uncertainty index of each parameter configuration region; the method comprises the steps of introducing an area expected improved REI index for better evaluating the optimization potential of an area, judging that the area has further exploration value and is kept or expanded when the optimization potential of the parameter configuration area is higher than a preset threshold, and judging that the optimization space of the area is limited and is contracted or switched when the optimization potential of the parameter configuration area is lower than the preset threshold.
  4. 4. The adaptive hybrid parameter optimization method for fusion of multitasking migration and regional potential modeling of claim 1 wherein REI is used to characterize the optimized potential benefit of a region as a whole defined as the average level of expected improvement values within the region: Wherein, the Represent the first A parameter configuration area is provided for the configuration of the parameters, Is a desired improvement value calculated based on the proxy model.
  5. 5. The adaptive hybrid parameter optimization method for fusion of multitasking migration and regional potential modeling of claim 1, wherein in the search behavior-based multitasking migration guidance step: The multi-task migration is based on parameter searching behaviors and region evolution characteristics, wherein the parameter searching behaviors at least comprise access sequence, sampling density distribution, region expansion or contraction frequency and performance convergence trend of parameter regions, the similarity between tasks at the parameter space structure and the optimization process level is evaluated by matching the early searching behaviors of the current tasks with the corresponding behavior characteristics of historical tasks, and when the similarity meets preset conditions, high-potential parameter region information or region-level agent models in the corresponding historical tasks are selected as priori and used for guiding region initialization, region priority ordering or sampling strategy setting of the current tasks, so that model multiplexing and migration guiding based on the searching behaviors is realized.
  6. 6. The adaptive hybrid parameter optimization method for fusing multitasking migration and regional potential modeling of claim 1, wherein in the adaptive updating step: Carrying out parameter sampling based on regional division of a parameter space, wherein the parameter space is divided into a plurality of parameter configuration areas, each parameter configuration area corresponds to one subspace in the parameter space, and historical performance statistical information and uncertainty measurement in the area are maintained; According to the optimization potential evaluation result of each parameter configuration area, preferentially selecting the parameter configuration area with higher potential for sampling, and generating a new parameter combination in the selected area based on an area-level agent model or an area constraint sampling strategy; When the continuous sampling result in the region shows improved performance or convergence trend change, the expansion, contraction or holding operation is dynamically executed on the corresponding parameter configuration region, so that region-level self-adaptive searching and resource allocation are realized.
  7. 7. The adaptive hybrid parameter optimization method for fusion of multitasking migration and regional potential modeling of claim 1, wherein the regional historical performance, the regional optimal performance and the regional potential indicators are combined to form a regional selection scoring function: ; Wherein, the The average performance of the region is indicated, Representing the historical optimal performance of the region, The potential of the region is represented and, Is a weight coefficient.
  8. 8. An adaptive hybrid parameter optimization system integrating multitasking migration and regional potential modeling is characterized by being applied to the adaptive hybrid parameter optimization method integrating multitasking migration and regional potential modeling as claimed in any one of claims 1-7, and comprising a performance monitoring and search track construction module, a hybrid parameter spatial regionalization modeling and potential evaluation module, a regional potential-based adaptive regional scheduling module, a search behavior-based multitasking migration guiding module and an adaptive updating module; The performance monitoring and searching track construction module is used for collecting performance feedback information corresponding to the current parameter configuration in the running process of a target program or a system, carrying out association storage on the performance feedback information and the corresponding parameter configuration combination, and constructing a parameter-performance historical data set and a parameter searching track for subsequent region modeling and multitasking migration analysis; The mixed parameter space regional modeling and potential evaluation module is used for carrying out unified representation modeling on a mixed parameter space containing continuous parameters and discrete parameters, dividing the parameter space into a plurality of parameter configuration areas, constructing an area-level proxy model based on historical performance and model uncertainty in each parameter configuration area, and calculating corresponding area optimization potential indexes; The self-adaptive region scheduling module based on region potential is used for adaptively executing region holding, expanding, shrinking or switching operation according to the optimization potential evaluation result of each parameter configuration region, and dynamically determining a target parameter configuration region of the current optimization stage so as to realize the balance between exploration and utilization; The multi-task migration guiding module based on the search behavior is used for analyzing parameter search behavior, region selection track and performance feedback distribution in the historical task and evaluating the similarity of the historical task and the current task in the search behavior level; The self-adaptive updating module is used for generating a new parameter combination in a selected parameter configuration area and applying the new parameter combination to a target program or system operation to acquire new performance feedback information, updating an area agent model, an area potential evaluation result and a search behavior characterization based on the new performance feedback information, and returning to execute the performance monitoring and search track construction step to form a continuous online self-adaptive optimization process.
  9. 9. An electronic device, the electronic device comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores computer program instructions executable by the at least one processor to enable the at least one processor to perform the adaptive hybrid parameter optimization method of fusion multitasking migration and regional potential modeling of any one of claims 1-7.
  10. 10. A computer readable storage medium storing a program, wherein the program, when executed by a processor, implements the adaptive hybrid parameter optimization method of fusion multitasking migration and regional potential modeling of any of claims 1-7.

Description

Adaptive hybrid parameter optimization method and device integrating multitask migration and regional potential modeling Technical Field The invention belongs to the technical field of automatic optimization and self-adaptive configuration management of a computer system, and particularly relates to a self-adaptive mixed parameter optimization method and device for fusing multitask migration and regional potential modeling. Background In recent years, with the continuous expansion of the scale of computer systems and the continuous enrichment of application scenes, the number of configuration parameters involved in the running process of software systems and computing platforms is continuously increased, and the interaction between parameter types and parameters is also increasingly complex. The system performance is often highly dependent on the reasonable configuration of the operation parameters, and the optimal parameter configuration of the system has significant differences under different application loads, different operation environments and different hardware platform conditions. Bolet et al state that in the OpenMP parallel optimization framework type commonly used for high-performance computing, if a suitable optimization strategy is adopted after a small amount of initialization sampling, the OpenMP parameter setting can bring about performance improvement of up to several tens of percent for some loads. For example, under the systematic sweep of OpenMP loop scheduling and thread count parameters in the implementation, certain loads can be reduced by about 20% -30% or more of run time by automatic optimization based on default policies. Therefore, how to realize automatic optimization of parameter configuration in the running process of a system has become an important research direction in the field of computer system performance optimization. In practical applications, the system configuration parameters generally include both continuous parameters and discrete parameters, such as thread number, cache size, scheduling policy, binding mode, and operation mode selection, so as to form a high-dimensional, hybrid-type parameter space. The parameter space is large in scale, and complex nonlinear coupling relation often exists between different parameters, so that the system performance shows high uncertainty on parameter change. The traditional parameter adjusting mode depending on manual experience or regular configuration is low in efficiency, and a stable optimization effect is difficult to obtain in a complex scene. In order to solve the above problems, a series of parameter optimization and automatic parameter adjustment methods have been proposed. One type of method is based on offline modeling or static search strategies, a performance model is built through a large amount of experimental data before the system is operated, and a parameter configuration scheme is determined according to the performance model. However, such approaches typically require high early test costs, and default system performance characteristics remain relatively stable during operation, and offline models tend to be difficult to continue to apply once the application load or operating environment changes. With the development of online optimization ideas, some research is beginning to focus on dynamically adjusting parameter configurations according to real-time performance feedback during system operation. The method realizes the on-line optimization of the parameters by continuously collecting the system operation indexes and updating the optimization strategy in the operation process. However, when facing a high-dimensional hybrid parameter space, a direct global search of the entire parameter space tends to result in inefficient sampling, slow convergence of the optimization process, and difficulty in achieving an ideal configuration within a limited time or resource budget. In response to the above problems, some studies have attempted to introduce probabilistic modeling, heuristic search, or uncertainty-based sampling strategies to improve the efficiency of parametric searching. For example, the potential performance of parameter combinations is predicted by building a proxy model, or trade-offs are made between exploration and utilization to reduce ineffective sampling. However, such methods still face a number of challenges in practical applications, on the one hand, the model construction and update costs rise significantly as the parameter dimensions increase, and on the other hand, the potential value differences exhibited by different parameter regions in the optimization process are not fully utilized, resulting in an imbalance in the allocation of computing resources. In addition, most existing online parameter optimization methods generally assume that the optimization process is developed for a single task or a fixed scenario, lacking in system utilization of historical optimization information. Whe