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

The scientific evolution allows the analysis of different phenomena with great accuracy and this results in a growing production of data to process. During last years, parallel computing has become a feasible solution, but a critical point is the efficiency in the I/O management. The characterization of I/O overhead is very important to improve the perfor- mance of parallel applications1, and the adoption of a parallel I/O in scientific applications is becoming a common practice. The image processing community has not completely appraised its use, although an increasing attention is paid to parallel computations. In this paper we address these issues in parallel image processing applications. We developed PIMA(GE)2 Lib, the Parallel IMAGE processing GEnoa Library; it provides a robust implementation of the most common image processing low level operations. During the design of the library, we look for a proper logical organization in the execution of I/O operations. We fixed the I/O pattern in the data access, and we made several tests about the logical organization in I/O operations to determine the most efficient strategy to apply in PIMA(GE)2 Lib. To achieve this goal we compared a master slave approach implemented with MPI 1, and a parallel I/O using the functionalities of the MPI-IO provided by MPI 2. In particular we use the ROMIO implementation2. In both cases we tested the interac- tion with the most common file systems for parallel and distributed applications, that are Parallel Virtual File System3, and Network File System4, both open source. Also aspects related with data layout on disk and efficient data distribution to parallel processes, for 3D images analysis tasks, are considered. This work represents an experimental study about the selection of a parallel I/O strat- egy. We show that MPI 2 parallel I/O, combined with PVFS 2, outperforms the other pos- sibilities, providing a reduction of the I/O cost for parallel image processing applications. Furthermore, at the best of our knowledge, PIMA(GE)2 Lib is one of the few examples of image processing library where a parallel I/O is strategy is adopted. 1. J. Dongarra, I. Foster, G. Fox, W. Gropp, K. Kennedy, L. Torczon, and A. White, The Sourcebook of Parallel Computing, Morgan Kaufmann, (2002). 2. ROMIO home page, http://www.mcs.anl.gov/romio 3. Parallel Virtual File System Version 2, http://www.pvfs.org 4. Sun Microsystems Inc., NFS: Network File System Version 3 Protocol Specification, Sun Microsystems Inc., Mountain View,CA, (1993).