Data Reshuffling in Support of Fast I/O for Distributed-Memory Machines

Achieving high-speed network I/O on distributedmemory systems is a hard problem because their architectures are, in general, ill-suited for communication with the external world. One of the problems is that messages are distributed over the private memories of the distributedmemory system. This can result in poor performance since communication includes a complex scatter/gather operation. This paper presents a strategy in which the task of creating large contiguous messages is performed on the distributed-memory system, thus minimizing the overhead on the network interface. The performance results for an implementation of this strategy for an iWarp system with a HIPPI interface board are presented. 1 Introduction Supercomputer applications have to communicate over a high-speed network, for example, to display results on a framebuffer, to read or write data from storage, or to interact with other computer systems as part of a distributed computing application. In the last few years, ne..