Parallel out-of-core sorting and fast accesses to disks

摘要

The paper addresses two problems. We investigate the problem of parallel external sorting in the context of a form of heterogeneous clusters then we investigate the impact of efficient disk remote accesses on the performance of external sorting. We explore three techniques to show how they can be deployed for clusters with proportional processor performances. We also validate the READ{sup}2 library, an efficient implementation of remote SCSI disk accesses. We derive a new parallel sorting algorithm that is adapted to the READ{sup}2 interface. The expected gain of using READ{sup}2 is compared to the measured gain for one external sorting implementation.