Research On Performance Of Different Programming Models Based Intel MIC Architecture

Intel MIC co-processor as a new heterogeneous computing platform,due to the characteristics that it is based on X86 architecture and supports Open MP,Open CL,p Thread and other parallel programming model,is widely used in high performance computing.Different parallel programming models have different performance for the specific application on Intel MIC platform.Choosing the efficient parallel programming model to improve the parallel efficiency as much as possible for the application has become a urgent problem.In this paper we parallelize non local means denoising algorithm and immuneodominance clone selection algorithm using Open MP and Open CL programming models on Intel MIC platform,respectively and compare the performance of the two parallel programming models on the Intel MIC platform.The main work of this paper is as follows:We analyze the serial non local means denoising algorithm and propose a parallel non local means denoising algorithm based on Open MP and Open CL on Intel MIC architecture.The denoising process of each pixel is executed on a thread of MIC core in the parallel algorithm.By comparing the PSNR of the serial and parallel algorithm,we prove the correctness and effectiveness of the parallel algorithm.We analyze the denoising performance and acceleration performance of two parallel algorithms with different search window size and similar window size.Then we analyze the performance of the Open MP-based parallel algorithm with different threads in MIC offload mode.In order to improve the performance of the Open MP-based parallel algorithm in MIC offload mode,we optimize the parallel algorithm with the SIMD instruction and compare the parallel performance of the optimized algorithm with the Open MP-based parallel algorithm in the MIC native mode and MIC offload mode,and get a speedup of more than 47 times.We study the performance of Open CL-based parallel algorithm with different trellis partition and analyze the time of the Open CL platform initialization and the kernel calculation,and get a speedup of more than 56 times.Finally,we compared the performance of the Open MP-based parallel algorithm and Open CL-based parallel algorithm on Intel MIC.For the problem that applying the immunodominance clonal selection algorithm to image feature selection is time-consuming,We propose parallel immunodominance clonal selection algorithm using Open MP and Open CL programming models by parallelizing all operators such as clone,crossover and mutation,affinity calculation and selection.The validity and correctness of the parallel algorithm are compared with the serial algorithm in three aspects: the classification accuracy rate,the feature subset dimension and the algorithm convergence trend.We compare the performance of the Open MP-based parallel algorithm with different threads in MIC offload mode.The time of Open CL platform initialization and the kernel calculation is analyzed respectively,finally the performance of Open MP-based parallel algorithm and Open CL-based parallel algorithm is compared on Intel MIC.