硬件归约的延迟模型和调度策略  

Delay model and scheduling policy of hardware-based reduction

作　　者：孙全宝[1] 张民选[1] 肖立权[1] 

机构地区：[1]国防科技大学计算机学院,湖南长沙410073

出　　处：《计算机工程与设计》2009年第8期2018-2021,2025,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(60633050)

摘　　要：通过硬件对归约进行支持是高性能互连网络发展的一种趋势。建立了无冲突和冲突条件下的归约延迟模型,分析了串行轮询、剩余报文集中和先到先服务3种调度策略对归约延迟的影响,表明了剩余报文集中策略能够取得多个归约操作在路由器上平均延迟的最小值。为了准确评测各种调度策略的性能,基于OMNeT++开发了微片级互连网络模拟器。模拟结果表明,先到先服务策略能够取得最优的性能,而在低维度互连网络中,串行轮询策略能取得良好的性价比。Supporting reduction at hardware level is a future trend of high performance interconnection networks. The reduction latency model with contention and non-contention is made. Then, the impact of three scheduling policies, round robin, residue concentration and first come first served, on the reduction latency is analyzed. To accurately evaluate the performance of different scheduling policies, a flit level network simulator based on OMNeT＋＋ is implemented. The simulation results show that first come first served can lead to highest performance and round robin can get better price to performance ratio in low radix networks.

关 键 词：并行处理 互连网络 聚合通信 归约 报文调度 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]