A HIGH SPEED COMMUNICATION MECHANISM FOR PARALLEL COMPUTING OVER ATM
Wu Li · Chinese Journal of Computers · 1998
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.