Performance Evaluation and Characterization of ….pptVIP

刘 莹 博士 副教授 yingliu@gucas.ac.cn 中国科学院研究生院信息科学与工程学院 ;简介;科研经历;科研经历;科研成果;提纲;数据挖掘;数据挖掘的动机—— 商业角度;数据挖掘的动机—— 科学计算角度;数据挖掘的起源;流程;数据挖掘的主要技术;聚类;异常点检测;分类;分类;预测;关联规则;顺序模式;时间序列;Why High Performance Data Mining?;Why High Performance Data Mining?;Progress in HPC - past 6 decades;TOP 10 Machines 7/2008;Date;Supercomputers in China;体系结构(Architectures);体系结构(Architectures);体系结构(Architectures);Parallel Programming;Parallel Programming;提纲;天体模拟（Cosmological Simulation）;;HOP Clustering Algorithm; Find the median particle on the longest axis Recursively bisect the particles along the longest axis Nearby particles are in the same sub-domain Each internal node contains boundary ; Generating density Traverse the tree to find Ndens neighbors for every particle Assign an estimated density to every particle Hopping Associate each particle with its densest neighbor Each particle “hops” to its densest neighbor till it reaches a particle that is its own densest neighbor Grouping Define particles associated to the same densest neighbor as a group Refine and merge groups; Key idea Load balance Assign approximate equal number of particles to each processor Minimize communication overheads Requests for potential required remote particles are packed into a single message, and the required particles are transferred to the requesting processors; Assign approximate equal number of particles to each processor Find the median particle in parallel on the longest axis Bisect particles along the longest axis Exchange particles between bisected processors Build local KD tree Maintain a global tree on each processor with no real particle transfer; Generate Density; Hopping Hop to its highest density neighbor Book the required remote particles and send out requests Transfer the required particles to requesting processors Grouping Particles linked to the same densest particle are defined as a group Refine gr