Research On Key Technologies Of Performance Data Mining And Algorithmic Optimization Based On Mobile Processor

More and more big data algorithms and tasks running in mobile computing platforms with limited resources.However,how to characterize the processor performance characteristics of these tasks running on mobile processors and how to optimize large data algorithms using processor performance characteristics are facing many challenges.The performance monitoring unit of the ARM mobile processor provides several hardware counters for counting specified micro-architecture events that occur at the clock cycle level during the operation of the processor or memory system.With the execution of the program,these hardware counters will produce a large amount of monitoring data,called Big Performance Data.Big performance data is widely used to describe task load,analyze program performance bottlenecks,and optimize micro-architecture.However,there are many challenges in describing processor performance characteristics of tasks running on mobile processors from big performance data and optimizing large data algorithms using it.The main challenges can be summarized as follows: 1)high data dimension.The dimension of large performance data is often as high as tens of thousands or even thousands.In the process of analyzing large performance data,it is usually based on experience to observe the data of some dimensions,or use the data of these dimensions to build statistical learning models.This method can not describe the task characteristics in an all-round way,and it is also difficult to find the relationship between the data.2)Big data information.The performance characteristics of large data tasks are different,and the event information obtained by hardware counter is too large to understand.Therefore,in this paper,we propose MobilePerfMiner,a performance data mining framework for ARM mobile processors.Using hardware counter,the performance model is constructed by using XGBoost algorithm iteratively,which ranks the importance of performance events for large data tasks and reduces the performance dimension of large data.Thus,the large data algorithm is optimized according to the performance characteristics described.The contributions of this paper are as follows:1.Performance data preprocessing.Integrate scattered data,read large data mosaic,normalization,structuring and other operations.2.Dimension reduction of performance data.The XGBoost algorithm is used to iteratively reduce the dimension of large performance data and get the feature importance ranking of hardware events.3.Characterization of processor performance.According to the sort of hardware events,important hardware events are obtained,and the processor performance characteristics of running large data tasks are described.In this paper,we use hardware counter to monitor 65 Performance Events supported by processor.Eighteen spark big data benchmark programs were tested.The performance characteristics of running spark benchmark program on mobile processor are analyzed.These performance characteristics are used to optimize the performance of the program.The experimental results show that: 1)the reduced dimension performance data can describe the performance characteristics more accurately,and the performance characteristics of large data tasks have generality and individuality.2)In the aspect of Spark optimization of micro-architecture events based on importance ranking,it saves 36% of program execution time by tuning Spark programs for instruction characteristics,37% of running time by memory system characteristics on average,and 21% of program execution time by tuning configuration parameters of data serialization.