Distributed heapmanagement using reference weights

Abstract

The progressive demand for computational power has stimulated the design of distributed memory multiprocessors. This in turn caused extensive research into efficient implementations of parallel functional, logic and object-based languages. Automatic heapmanagement using garbage collection is often an integrated part of those implementations. However, distributed garbage collectors can introduce a bottleneck for distributed processing. Both the number of communication messages and the processor idle time may increase untolerably. Independently, Bevan and Watson [Be87, Wa87] proposed a method which solves these problems at the cost of using pointer indirection cells. This introduces an unacceptable delay to every object access. In this paper an alternative is described which does not suffer from this access delay. Besides explaining our method and showing its additional advantages, it is demonstrated that within the framework of pipelined RISC architectures, efficient architectural support for this method is easily achievable.

This research has been supported by IBM, as a part of a joint study agreement.