Design and Implementation of a Parallel File Subsystem on TreadMarks

Abstract

In TreadMarks and most current Distributed Shared Memory systems, file accesses are usually handled sequentially by one node. A large amount of network traffic is generated between this node and the other nodes to distribute the input data and to collect the resultant data. To reduce network traffic and shorten file access time, we have dedeloped a parallel file subsystem on TreadMarks. A parallel file is partitioned and distributed among all nodes to parallelize file accesses. With our variable data distribution scheme, the network traffic for file accesses is greatly reduced. The total execution time of a 2000*1000 20-iteration Successive Over Relaxation (SOR)program on 8 nodes is reduced from 95 seconds with sequential I/O to41 seconds with parallel I/O, which that of 1024*1024 Matrix Multi-plication on 8 nodes is reduced from 258 seconds with sequential I/O to 236 seconds with parallel I/o.