Parallel I/O support for HPF on clusters

Clusters of workstations are a popular alternative to integrated parallel systems designed and built by a vendor Besides their huge cumulative processing power, they also provide a large data storage capacity which allows efficient implementations of large-scale applications which are I/O-intensive. This paper proposes a language, compiler and runtime software solution to the problem of parallel I/O on clusters. The proposal is presented in the context of High Performance Fortran (HPF) and its compilation and runtime environment that is being developed at the University of Vienna. The system provides efficient support for explicit I/O operations on parallel files, accesses to sections of multi-dimensional arrays stored in parallel files, checkpoint/restart operations, and time-step input and output operations. The paper also presents experimental performance results using the implementation of the developed software on a Beowulf-class cluster system