Exploiting Single-Assignment Properties to Optimize Message-Passing Programs by Code Transformations

The message-passing paradigm is now widely accepted and used mainly for inter-process communication in distributed memory parallel systems. However, one of its disadvantages is the high cost associated with the data exchange. Therefore, in this paper, we describe a message-passing optimization technique based on the exploitation of single-assignment and constant information properties to reduce the number of communications. Similar to the more general partial evaluation approach, technique evaluates local and remote memory operations when only part of the input is known or available; it further specializes the program with respect to the input data. It is applied to the programs which use a distributed single-assignment memory system. Experimental results show a considerable speedup in programs running in computer systems with slow interconnection networks. We also show that single assignment memory systems can have better network latency tolerance and the overhead introduced by its ma...

• Type: Conference
• Year: 2004
• Where: IFL
• Authors: Alfredo Cristóbal-Salas, Andrey Chernykh, Edelmira Rodríguez-Alcantar, Jean-Luc Gaudiot
• Keywords: Distributed Memory Parallel, IFL 2004, Memory Systems, Message-passing Optimization Technique