ELSEVIER Journal of Systems Architecture 44 (1998) 241-260 JOURNAL OF SYSTEMS ARCHITECTURE ParaStation: Efficient parallel computing by clustering workstations: Design and evaluation Thomas M. Warschko .,1, Joachim M. Blum, Walter F. Tichy O ./ ' " Univcrsilv o/Karlsruhe. Dept. ,/ p~/ormatwsAm Fasanengarten 5, D-76128 Karlsruhe, Germany Received 1 November 1996: received in revised form 3 February 1997: accepted 1 April 1997 Abstract ParaStation is a communications fabric for connecting off-the-shelf workstations into a supercomputer. The fabric employs technology used in massively parallel machines and scales up to 4096 nodes. ParaStation's user-level message passing software preserves the low latency of the fabric by taking the operating system out of the communication path, while still providing full protection in a multiprogramming environment. The programming interface presented by ParaStation consists of a UNIX socket emulation and widely used parallel programming environments such as PVM, P4, and MPI. Implementations of ParaStation using various platforms, such as Digital's AlpbaGeneration work- stations and Linux PCs, achieve end-to-end (process-to-process) latencies as low as 2 ~s and a sustained bandwidth of up to 15 Mbyte/s per channel, even with small packets. Benchmarks using PVM on ParaStation demonstrate real ap- plication performance of 1 GFLOP on an 8-node cluster. Keywordw Workstation cluster; Parallel and distributed computing; User-level communication; High-speed interconnects 1. Introduction Networks of workstations and PCs offer a cost- effective and scalable alternative to monolithic * Corresponding author. E-mail: warschko@ira.uka.de. i WWW: http://wwwipd.ira.uka.de/parastation. supercomputers. Thus, bundling together a cluster of workstations - either single-processors or small multiprocessors into a parallel system would seem to be a straightforward solution for computa- tional tasks that are too large for a single machine. However, conventional communication mechan- isms and protocols yield communication latencies that prohibit any but very large grain parallelism. 1383-7621/0165-6074198/$19.00 © 1998 Elsevier Science B.V. All rights reserved. PIISl383-7621(97)00039-8