High Performance Parallel Computing with ParaStation3

Abstract

The key to efficient and wide spread parallel computing on workstations clusters is a high speed communication subsystem with well known programming interfaces and a cluster management system which makes the use of a cluster as simple as for a single machine. ParaStation3 fullfills these requeiremnts. It is a high speed communication subsystem for clusters of workstations in combination with Operating System enhancements to view the whole cluster as one entity. It offers all kind of programming interfaces, such as MPI, PVM, Sockets, and Java RMI close to the hardware limits. It achieves latencies as low as 7.5 µs and throughput as high as 350 MByte/s (biderectional data exchange). The Operating System enhancements for the single system view include fault tolerance, transparent node and process management, and load balancing. The system is implemented on a variety of platforms (Alpha workstations running Tru64 Unix, as well as Intel and Alpha"s running Linux) and is proven to operate large clusters.