In this paper we have carried out an experimental study of the I/Olatency of a distributed file system over a parallel multimediaplatform. Our experimental platform uses NFS (Network File System) tosupport an application characterized by its high processing and storagerequirements. We evaluate the overall system by varying several NFSparameters as well as the underlying network topology and data accesspatterns. We have also examined the relevance of two other importantfile management access mechanisms: caching and prefetching. Ourexperimental results have allowed us to identify the major systemparameters affecting the performance of NFS. We then propose the use ofa predictive data access mechanisms. Under this mechanism, the fileserver should detect the similarities among the most recent I/O requestsincoming from a set of cooperating clients. Once having identified theaccess pattern, the file server will be able to multicast the requireddata even before receiving explicit I/O data requests from the clients.Under this predictive mechanism, the system performance should improveas the number of explicit I/O requests directed to the file serverdecreases
展开▼