异构集群上的宏基因组基因聚类优化-CenterforHPC-上海交通大学.PDFVIP

计 算 机 研 究 与 发 展 DOI:10.7544/issn1000-1239.2015.******** Journal of Computer Research and Development 卷(期):起止页,年 异构集群上的宏基因组基因聚类优化 1 1 2 1,3 1 韦建文 许志耿 王丙强 Simon See 林新华 1 上海交通大学高性能计算中心，上海 200240 2 国家超算深圳中心，深圳 518055 3 NVIDIA ，新加坡 (james@sjtu.edu .cn) Accelerating Gene Clustering on Heterogeneous Clusters Jianwen Wei1 and Zhigeng Xu1 and Bingqiang Wang2 and Simon See1,3 and James Lin1 1(Center for High Performance Computing, Shanghai Jiao Tong University, Shanghai 200240) 2 (National Supercomputing Center in Shenzhen, Shenzhen, 518005) 3 (NVIDIA, Singapore) Abstract Metagenome clustering is a novel approach to detect flaw genes which relies on massive gene data, effective clustering algorithms and efficient implementation. In clustering, calculating correlation matrix is essential, accounting most of compute time. To take a gene repo as an example, which has 1300 samples and million genes, it will take about 27 years to cluster them. Therefore, developing efficient implementations for calculating correlation matrix is most essential. After analyzing the algorithms, we propose and take several optimization approaches. First, we implement an efficient multithread one using OpenMP dynamic scheduling. Secondly, we further improve the performance by utilizing cache on CPU and shared memory on GPU efficiently. Thirdly, we implement a naïve loadbalance work distribution which works well on the MPI program on CPU. Compared to the unoptimized single-threaded CPU program, the two fasted one, MPI+OpenMP on 256 CPU cores and MPI+CUDA on 6 GPU cards, achieve 238.8 and 263.8 speedups. Key words gene clustering; heterogeneous computing; cache optimizatio