LOGO
周兵兵学术报告通知
时间:2017-11-16 08:44:54
时间:2017年11月16日9:00
地点:华中科技大学东五楼210
报告题目和内容:一种面向大规模高性能集群上成对计算的并行算法
Abstract – All pairwise computation is defined as performing computation between each and every pair of the elements in a given data set. It is often a necessary first step in a number of bioinformatics applications. Many of such applications require a number of terabytes of main memory and take multiple peta floating point operations to complete the computation. Therefore, large HPC clusters may be needed to tackle these large-scale computational problems. Currently, most parallel algorithms for all pairwise computation used in bioinformatics applications are designed and implemented using master/worker programming paradigm. Though easy to be implemented on MapReduce and Apache Spark computing platforms, this type of parallel algorithms for all pairwise computation inevitably requires the sequences to be loaded multiple times from I/O disks to compute worker nodes. To solve large-scale computational problems using a large number of compute nodes, the disk I/O cost can become prohibitively high. Thus these parallel algorithms are neither efficient, nor scalable for large-scale all pairwise computation on large HPC clusters. Another type of parallel algorithms for all pairwise computation is developed using SPMD (Single Program and Multiple Data) programming paradigm with explicit message passing for communication between the compute nodes. These parallel algorithms load the data set only once from the I/O disk and make effective use of fast interconnection networks in HPC clusters for parallel communication between compute nodes. Thus they are more efficient and scalable for large HPC clusters. In this talk I introduce a new parallel algorithm for all pairwise computation. A prominent feature of this algorithm is that the communication cost between compute nodes is only half the cost of the most efficient algorithm available so far. Experiments on a Cray XC40 HPC supercomputer show that this new algorithm is more efficient and scalable for large-scale all pairwise computation on large HPC clusters.
 
周兵兵教授简历:周兵兵教授在悉尼大学信息技术学院任职。主要研究领域是分布式与并行计算,算法设计与分析,云计算以及生物信息。在IEEE Trans on computers、IEEE TPDS等期刊发表论文40余篇,在GLOBECOM、HPDC等国际会议上发表论文60余篇。是IEEE Trans on computers science,TPDC,Journal of PDC等期刊的审稿人。
 更多 更多 更多 更多