Parallel Compression of Correlated Files

摘要

Economy-based admission control of jobs in a grid, or migration of guest jobs from a disconnecting cluster in a grid, as well as checkpointing parallel jobs in a cluster to a central repository are demanding tasks that can exhaust essential resources such as the communication networks, due to the requirement to quickly move large amounts of data from many nodes. Compressing memory images might make these operations more efficient provided that the overall throughput is increased. Existing serial compression algorithms are not suitable for such purposes because they do not exploit interfile redundancy. This paper presents decentralized algorithms for parallel compression of correlated memory images of a job in a cluster or in a grid. The algorithms use block suppression to eliminate inter-file redundancy. They take advantage of the multiple processor environment to simultaneously map memory blocks to hash values in order to detect redundancies. It is shown that exploiting interfile redundancy of correlated files can increase the overall transfer throughput of parallel jobs. It is also shown that combining serial compression with our algorithms further increases this throughput. The paper presents the algorithms and their performance.