摘要
高性能计算机系统越来越多采用集群系统,集群系统的性能极大地依赖于通信接口。基于片外SRAM保存地址变换表的用户级通信方法,极大地增加了芯片和系统的设计复杂度和成本。在传统基于I/O总线的HCA基础上,提出并实现了一种新型的Free-Memory的众核处理器片上通信接口,去掉了本地存储器接口,通过高效的cache管理策略降低地址变换开销。测试结果表明我们实现的通信接口在降低了芯片和系统实现复杂度和成本的同时,还获得了比Infiniband的QDR HCA更好的通信带宽和延迟。
High-performance computing has witnessed a tremendous growth and acceptance over the last decade,primarily due to the availability of clusters.The performance of these clusters hinges upon the communication interface.User level communication based on storing address translation table in off-chip SRAM has deeply increased the design complex and cost of chip and system.The paper put forward and implemented a novel Free-Memory communication interface of many cores processor which differs from traditional HCA based on I/O bus,without local memory interface and reduced cost of address translation by efficient cache management method.Experimental results show the communication interface which we implemented not only can reduces the design complex and cost of chip and system,but also can achieve better bandwidth and latency than infiniband QDR HCA.
出处
《国防科技大学学报》
EI
CAS
CSCD
北大核心
2011年第2期145-149,共5页
Journal of National University of Defense Technology
基金
核高基重大专项--面向科学计算的多核CPU资助项目(2009ZX01028-002-002)