摘要： 面向集群系统的通信故障,研究了如何在消息传递层采用故障接管实现通信子系统的透明容错,并描述了基于高性能通信接口NICHAL的容错MPI(R-MPI)实现,测试数据表明该实现有效利用了RDMA特征实现容错通信协议。

关键词: 集群, MPI, 故障接管, 容错

Abstract: Fault-tolerent message passing middleware is one of the key techniques for parallel computing in large scale clusters.This paper presents the authors' work on R-MPI,a high speed NI based MPI implementation,which introduces failover mechanism to recover from disconnection of network.

中图分类号: 

• TP393