Research On Preprocess-based Parallel Algorithm Optimization On Hybrid Structure Platform And Implementation

Currently, hybrid structure platform composed of CPU,GPU,FPGA and anotherdifferent computer components is one of the hot spots of research for high performancecomputing. How to improve utilization of computing power and resources on hybridstructure platform is a challenging task.Nowadays, studies on hybrid structure platform mainly concentrate on the schedulingpolicy. Though scheduling can perform the characteristics of hybrid structure platform, andeffectively reduce the computational time of application. However, just relying onscheduling to complete the tasks on hybrid structure platform has its limits. Becausescheduling often relies on the original program structure, and the computing power onhybrid structure platform is restricted.Thus, we put forward a parallel optimization scheme based on preprocess. The purposeof preprocess is to adjust the structure of the original program which makes it more suitablefor hybrid structure platform to compute. Combined with scheduling policy, we can furtherexploit the computing power of hybrid structure platform, and effectively reduce thecomputation time of application. We select two typical applications(computer simulation ofElectrocardiogram and support vector machine) for validation.In this paper, the main work is as follows:First of all, we have completed the algorithm design and verification scheme based onpreprocess. The content of preprocess algorithm design mainly concentrated on two parts.One is the load prediction method, and the other is the pre-calculation method. We mainlyuse these two methods to deal with the calculation problems on hybrid structure platform.The main process of validation protocols for preprocess method has four aspects: theanalysis of application, preprocess, scheduling, and computing on hybrid structure platform.Then based on the design idea of preprocess, we complete the corresponding algorithm design and implementation for two applications, and carry on the analysis of the experimentresults. The parallel algorithm optimization on computer simulation of Electrocardiogramis mainly based on the load prediction method. The adjusted program structure of computersimulation of Electrocardiogram is more advantageous for hybrid structure platform toexploit its performance which can effectively reduce the computation time of the program.The optimization of support vector machine (SVM) algorithm mainly adopts thepre-calculation method. Through the preprocess work of two pre-calculations combined withthe characteristics of hybrid structure platform, program running time is greatly reduced.