基于信令寻径式网络的高性能容错VIA设计与实现

Design and Implementation of High Performance Fault-Tolerant VIA Based on Token Route Network

VIA为机群系统定义了高带宽、低延迟的通信机制。基于交换技术的信令寻径式网络采用通信信令进行通信支持,能很好的解决高性能计算中数据突发通信问题。本文在信令寻径式网络中实现了VIA通信接口TR-VIA(TokenRoute-VIA)。TR-VIA采用软硬件协同设计思想,来优化VIA的设计;采用发送/接收代理机制和容错硬件通信协议,实现了高性能的容错VIA通信接口;采用公共通道实现了低延迟数据通信的目标。通过性能的分析比较,TR-VIA通信延迟在16字节时达到5μs,带宽效率达到96.96%,与目前其它VIA实现技术相比,TR-VIA具有很大的优越性。

Virtual interface architecture (VIA) established a communication model with low latency and high bandwidth in cluster system. TR-NET (Token Route Network) based on switch technique adopts communication token to support data transferring. And it can resist heavy load impact in cluster computing successfully. The paper presents a implement of VIA on TR-NET named TR-VIA. The idea of software-hardware co-design is used to optimize the designing of TR-VIA; the high performance fault-tolerant VIA is designed by the technique of the communication agent and the fault-tolerant hardware communication protocol; the page-tables in user layer decrease the overhead in virt-to-phy translation; the public channel decrease the communication latency; Synchronization token realize the high bandwidth efficiency. The experimental results indicate that the bandwidth efficiency is 96.96%, the lowest one-way latency is 5μs, which show that the performance of TR-VIA surpasses that of other VIA in some aspect.