PNNL

Abstract

This paper describes a new methodology for implementing a common set of collective message-passing operations on clusters based on symmetric multiprocessor nodes. Called Shared-Remote-Memory (SRM) collectives, this approach replaces the point-to-point message passing, used traditionally in collective message-passing operations, with a combination of shared and remote memory protocols. The approach reduces the amount of data movement and avoids the overheads associated with tag matching and buffer management complexities in implementing the general-purpose point-to-point message-passing protocols. Experimental results obtained on the IBM SP with 16-way nodes show that the SRM collectives technique outperforms the highly optimized IBM implementation of MPI across a wide range of message sizes and processor counts.