A Java Implementation of MPI-I/O-Oriented Random Access File Class in AgentTeamwork Grid Computing Middleware

Abstract

Abstract-MPI-I/O defines a high-level file interface that enables multiple ranks to share a random access file. It would be highly attractive to grid-computing users that files are automatically partitioned and transfered to remote sites where their jobs can access the files through MPI-I/O. We are currently implementing in the AgentTeamwork grid-computing middleware system a series of these file-handling features including: file partitioning into strides, stride distribution to multiple processes, stride access through our MPI-I/O-oriented random access file class, stride exchange among processes, and barrier synchronization support. Particularly focusing on a design of our random access file class, this paper presents an implementation and performance results of AgentTeamwork's file-partitioning and stride-maintenance schemes.

An implementation of parallel file distribution in an agent hierarchy

“…Its obvious drawback is the more ranks the more repetitive scans of the same file. An improvement is to limit this file-scanning work to only once by reading an input file with a byte basis, checking which ranks need each byte, and writing it to the corresponding rank-based file stripes [11]. However, this scheme yet suffers from another drawback that cannot perform bulk read/write operations to a file, thus increasing OS interventions.…”