Task-Migration Support for Parallel Computing on Shared Virtual Memory Clusters

摘要

Shared virtual memory (SVM) is an attractive approach for providing simple parallel programming environments on a cluster of personal computers or workstations, and symmetric-multiprocessor clusters. It permits programmers to assume the existence of shared memory image across physically distributed-memory systems, obviating the need of explicit inter-process communication operations. While several memory coherence protocols for improving SVM performance have been proposed and studied extensively, very little attention has been given to improve its reliability despite the fact that wider acceptance of SVM in production systems is depending on it. In this paper, we propose a protocol suite for task-migration support in SVM. Our protocol suite consists of a checkpointing protocol, a message-logging protocol, and a fast reconfiguration protocol. Our experimental results demonstrated that our checkpointing protocol incurs the execution time overhead of 0.01% to 4% when task-migration is initiated every 2-minute interval. Our message-logging protocol incurs only 1% to 3% overhead in logging messages and enables our reconfiguration protocol to drastically accelerate the migration process, ranging from 2% to 67% by speeding up the execution replay at the target node. Hence, our protocol suite makes task-migration support in SVM an attractive reality.