READ2: Put disks at network level

Grand challenge applications have to process largeamounts of data, and then require high performance IOsystems.Cluster computing is a good alternative to proprietarysystem for building cost effective IO intensive platform:some cluster architectures won sorting benchmark(MinuteSort, Datamation)! Recent advances in IO componenttechnologies (disk, controller and network) let us expecthigher IO performance for data intensive applicationson cluster.The counterpart of this evolution is that muchstress is put on the different buses (memory, IO) of eachnode which cannot be scaled.In this paper we investigate astrategy we called READ2 (Remote Efficient Access to DistantDevice) to reduce the stress.With READ2 any clusternode accesses directly to remote disk: the remote processorand the remote memory are removed from the control anddata path: Inputs/Outputs don't interfere with the host processorand the host memory activity.With READ2 strategy,a cluster can be considered as a shared disk architectureinstead of a shared nothing one.This papers describes animplementation of READ2 on Myrinet Networks.First experimentalresults show IO performance improvement.