Parallel I/O aspects in PIMA(GE)2 Library

Input/Output operations represent a critical point in parallel computations. and often results in a bottleneck with a consequent reduction of application performance. This paper describes the parallel I/O strategy applied in PIMA(GE)(2) Lib, the Parallel IMAGE processing GEnoa Library. The adoption of a parallel I/O results in an improvement of the performance of an age processing applications developed using the library. To achieve this goal we performed an experimental study, comparing an MP 1 implementation of a classical master-slave approach, and an MPI 2 implementation of parallel I/O. In both cases we considered the use of two different file systems. namely NFS and PVFS. We show that MPI 2 parallel I/O. combined with PVFS 2, outperforms the other possibilities, providing a reduction of the I/O cost for parallel image processing applications, if a suitable programming Paradigm for I/O organization is adopted.