ParFiSys
The continuous growth of the MPP computing power, which is expected to increase in the next years, has not been corresponded with an equivalent improvement on I/O technology. As a result, computation, memory and I/O are increasingly unbalanced every year, with more MPP applications being limited by I/O bandwidth. While a lot of research effort has been devoted during the past years to increase MPP computation capacity and to solve MPP memory problems, an smaller effort has been devoted during the past years to solve the I/O problem, yielding to the what has been defined as the I/O crisis. We propose to use parallel file systems, as ParFiSys, to alleviate this problem as it seems to be a good approach to take advantage of the inherent parallelism of distributed I/O systems. ParFiSys is a parallel file system developed at the Universidad Politécnica de Madrid (UPM), Spain. The main goals of ParFiSys are to provide I/O services to scientific applications requiring high I/O bandwidth, to minimize application porting effort, and to exploit the parallelism inherent to generic message-passing multicomputers, including processing nodes (PN) and I/O nodes (ION). The first version of PaFiSys was developed within the General Purpose MIMD (GPMIMD) project (P-5404) of the CEE ESPRIT program. ParfiSys was named Cache Coherence File System (CCFS) in that project, and this name is used in many documents shown here. ParFiSys is being used to explore a variety of data distribution, distributed caching and prefetching strategies: disk data striping factors, file and disk block prefetch policies, caching policies, cache coherence models, and I/O patterns.