推测执行技术在HPMR系统通信优化中的应用  

Application of speculative execution technology for communication optimization in HPMR

作　　者：王昊[1] 王向前[2,3] 郑启龙[2,3] 

机构地区：[1]中国电子科技集团公司第三十八研究所,安徽合肥230088 [2]中国科学技术大学计算机科学与技术学院,安徽合肥230027 [3]安徽省高性能计算重点实验室,安徽合肥230026

出　　处：《中国科学技术大学学报》2010年第11期1191-1196,共6页JUSTC

基　　金：国家自然科学基金重点项目(60533020);核高基重大专项(2009ZX01034-001-001-002);安徽省自然科学基金(090412068)资助

摘　　要：HPMR(high performance MapReduce)系统是一个采用MapReduce模型的高性能计算软件支撑平台.在进行高性能计算时,HPMR往往需要很多轮Map-Reduce过程,中间的通信过程由KV路由和KV传输组成.KV传输依照KV路由阶段产生的KV路由表进行.HPMR程序的典型通信特征是每一轮的KV传输以高概率使用上一轮的KV路由表.根据这一特点,提出在HPMR的通信模型中引入推测执行技术,由此减少了KV路由的次数,使得HPMR的通信性能得到大大地提升.HPMR （high performance MapReduce）system is a high performance computing software supporting platform.It is common for HPMR to do many rounds of the Map-Reduce process in high performance computing.The middle communication process consists of KV routing and KV transfer.KV transfer is carried out with the help of the KV table generated by KV routing.The notable feature of communication in HPMR program is that each round of KV transfer adopts the KV table of the last round with high probability.In view of to this characteristic,speculation execution technology is introduced into the communication model of HPMR,which reduces the number of KV routings and greatly enhances communication performance of HPMR.

关 键 词：MAPREDUCE HPMR 推测执行 KV路由 KV传输 

分 类 号：TP338.6[自动化与计算机技术—计算机系统结构]