Emmanuel Cecchet
Alexandre Meyer
Alexandre.Meyer@imag.fr
INRIA
In this paper we compare the efficiency of a supercomputer with
hardware shared memory with that of a cluster of workstations using
software Distributed Shared Memory (DSM). The difference in performance
is studied by running ray tracing applications on both architectures.
We have ported Stingray, a parallel cone tracer developed on a SGI
Origin 2000 supercomputer, on a cluster interconnected by a Scalable
Coherent Interface (SCI) network and a software DSM called SciFS. We
present the implementation issues and compare the results obtained with
each architecture. Further we discuss the tradeoff -
price/performance/programming ease - of both architectures. We show
using Stingray that a modest 12 nodes SCI cluster with an efficient
software DSM is 5 times cheaper and can perform up to 2.3 times better
than a SGI Origin 2000 with 6 processors. We therefore consider that
software DSM is well suited for this kind of application and provides
both ease in programming and scalable performance.