An Eecient Communication System for Parallel Computing on Networks of Workstations

This paper presents the design and implementation of an e cient communication system, Pupa, developed speci cally for supporting parallel computing on networks of workstations (NOW) architecture. The emphasis of Pupa is on both performance and service functionality. Pupa provides a signi cantly lower latency communication support compared to TCP/IP and is as powerful as the latter as an underlying communication system for parallel computing in a NOW environment. Moreover, it can coexist with TCP/IP without causing any performance degradation. The key architectural ideas of Pupa are { a simpli ed FIFO based bu er management scheme, an optimistic ow control mechanism, the decoupling of reliability related operations from the critical path of message delivery, and doing away with unnecessary functionalities such as checksumming. We have the rst Pupa prototype operational and we have successfully ported a communication library and tested a set of parallel applications on a network of 486-DX2 ISA PCs connected by a 10Mbits/s Ethernet. Programs when running on top of the Pupa system run upto 60% faster than when running on top of TCP/IP in this environment. As a measure of the raw performance improvement, we have observed the message latencies in the Pupa system in the best case to be less than one-fourth of the TCP/IP latencies.