Improving DSZOOM's Run Time System

Abstract

DSZOOM is a software-based distributed shared memory (DSM) system intended to run on a cluster of computers that communicates through interconnect hardware. In order to run on a multi-system image (MSI) cluster, meaning a cluster where each node has a separate operating system running, there is a need for a run-time system that supports process distribution. In this thesis, we have implemented a run-time system that will distribute processes across the cluster without the assistance of the operating system. This is done using the robust and widely accepted message passing interface (MPI) standard. The run-time system is also optimized to take advantage of UltraSPARC's Block Load/Store instructions. We have also implemented hierarchical lock technology in DSZOOM. The lock algorithm creates node affinity by keeping the lock in a node for some time if there is contention. The micro-benchmark study shows a 500% speed-up on a two node cluster with 20 processors. The hierarchical lock can run on any number of nodes.

Available as PDF (340 kB)

BibTeX file entry: Ekstrom:2004:jan