Parallel cluster architectures are an effective way to build high-performan
ce computer systems at a moderate price. Consequently, clusters will become
very attractive to parallel applications and high-end parallel servers: wh
ich are usually built around distributed object middleware. Commonly used m
iddleware layers like CORBA have been specifically designed for heterogeneo
us computing within relatively slow networks. In contrast PC and workstatio
n clusters are homogeneous systems, which can be connected by very high-spe
ed user-level networks. Only very lean and efficient middleware layers are
therefore able to use such kinds of communication infrastructure effectivel
y. In this paper we describe the architecture and implementation of ARTS, a
high-performance middleware layer developed for the PEACE family of parall
el operating systems, and show that this middleware is even applicable to t
he implementation of extremely performance-sensitive virtual shared memory
(VSM) systems. (C) 1999 Academic Press.