Improving Cluster IO Performance with READ^2

Abstract

(eng) Grand challenge applications often need to process large amounts of data so high performance IO systems are needed. Cluster computing is a good approach to build cost effective IO intensive platform: sort benchmark records (MinuteSort, Datamation) are held by cluster architecture! New advances in IO technologies (disk, controller and network) let expected significant performance improvements for data intensive applications on cluster architectures. The counterpart of this evolution is that more stress is put on the different shared buses of each node. In this paper we investigate a solution to reduce the bus pressures to improve IO performance we called READ^2 (Remote Efficient Access to Distant Device). With READ^2, a cluster node is able to directly access to a remote disk without interfering with the remote processor and remote memory. With READ^2, a cluster can be considered as a \emph{shared disk} architecture instead of a \emph{shared nothing} one, and inherits works from the SAN community. This papers presents what are the architectural benefits of READ^2, i.e. a better use of IO and memory buses which may improve performance of streaming application.