A Scalable Distributed Shared Memory

Abstract

Parallel computers of the future will require a memory model which offers a global address space to the programmer, while performing equally well under various system configurations. We present a logically shared and physically distributed memory to match both requirements. This paper introduces the memory system used in the ADAM coarse-grain dataflow machine which preserves scalability by tolerating latency and offers programmability through its object-based structure. We show how to support data objects of arbitrary size and different access bandwidth and latency characteristics, and present a possible implementation of this model. The proposed system is evaluated by analysis of the bandwidth and latency characteristics of the three different object classes and by examination of the impact of different network topologies. Finally, we present a number of simulation results which confirm the previous analysis.