一种基于ATM的支持并行处理的高速通信机制

A HIGH SPEED COMMUNICATION MECHANISM FOR PARALLEL COMPUTING OVER ATM

随着高速网络（如ATM）的发展以及工作站性能的不断提高，工作站网络（NOW）作为～种新型的并行计算结构越来越受到人们的重视．传统的传输协议和报文传递系统不能充分利用高速网的传输能力．本文提出一种基于ATM的支持并行处理的高速通信机制HPMPA．在HPMPA中，可靠的端一端传输协议HSTP为并行应用提供高速可靠的数据传输，而不可靠的端-端传输协议UTP则提供不可靠的高速数据报服务，以混合树结构（Hybrid-tree）为基础的各种群通信操作则为并行应用提供各种高性能的群通信服务．

With the wide availability and the rapid development of high speed network such as ATM, Myrinet, NOW(network of workstations) plays an important role in parallel computing. Traditional transport protocol such as TCP and message passing system can not take advantage of the capacity of high speed networks fully. This paper describes a high speed communication mechanism for parallel computing over ATM, called HPMPA, which consists of five layers: ATM AAL(adaptation layer), LLCL (low level communication layer), HLCL (high level communication layer), GM (group management), and AL (application layer). ATM AAL provides sequenced, unduplicated, but unreliable data transport services. LLCL provides reliable and unreliable point-to-point transport services. In LLCL, a new transport protocol, HSTP(high speed transport protocol), is put forward to provide low latency data transport services for applications on NOW by selective repeat,periodic states exchange,hybrid acknowledgments,block transport for bulk data, and so on.Another transport protocol in LLCL, UTP (Unreliable Transport Protocol), provides unreliable end-end transport services. Collective Communications services areprovided by HLCL, which are implemented based on a new kind of collective communication structure, hybrid-tree which fully takes advantage of hardware multicast provided by ATM switches. GM implements the management of process groups in parallel computing. The parallel applications in AL will achieve good performance by accessing above communications services.