Home-based SVM Protocols for SMP Clusters: Design, Implementation and Performance

Abstract:

Shared virtual memory has been used as a cost-effective means of
providing the attractive shared-memory model to programmers. In the
past, most SVM implementations have been focussed on clusters of
uniprocessor systems. Due to the increasing popularity and declining
costs of small scale shared memory multiprocessor systems (SMPs), they
have become attractive candidates for replacing the uniprocessor nodes
when building a shared virtual memory layer. At present, there is no
definite answer on what is the best approach to implementing an SVM
layer for a cluster of SMPs, whether such a protocol is necessarily
different from one designed for uniprocessor nodes, and what accepts
of this difference actually results in increases in performance.

To answer these questions and explore the design space of SVM on SMPs,
we have implemented a home-based lazy release consistency (HLRC)
protocol on a cluster of 4-way Intel PentiumPro SMPs interconnected
with Myrinet. This paper presents numerous protocol design options we
were faced with and their implications. We are able to achieve very
promising performance on some applications investigated so far. We
have investigated (quantitatively) the tradeoffs in our system to
validate some of the choices we have made. The tradeoff investigation
is ongoing, and we will complete it for the final paper.