Low Overhead Object Communication Scheme in CMP Implementation of Object-Oriented Programming

Abstract

Object-oriented (OO) programming appears as a promising framework for parallel programming. Objects which are inherently concurrent communicate through message passing. But most of OO languages choose to avoid supporting message passing directly because the implementation is too complicated. As we enter the era of CMP with multiple cores on a single die, real message passing can be realized. Objects can be allocated to cores and communicate through the network on chip. In this paper, we propose a hardware object communication scheme based on CMP, which realizes the message passing with low overhead and high efficiency. The locality of messages in object-oriented programs is discussed particularly. We implement our scheme in SystemC and evaluated it using a SystemC and ISS co-simulation method. The results show the hit ratio of message mapping can reach 92% and the speedup of parallel OO programs executed in 9-core CMP is about 6 compared with sequential OO programs at best.