Active Memory Clusters: Efficient Multiprocessing on Commodity Clusters

We show how key insights from our research into active memory systems, coupled with emerging trends in commodity network technology, are leading toward the realization of hardware distributed shared memory (DSM) on clusters of industry-standard workstations. We call the result of this convergence active memory clusters. After discussing the current state of the art in hardware DSM, clusters, and software DSM architectures, we highlight the key differences between hardware and software DSM systems and show how these differences are rapidly disappearing in commodity systems—with the notable exception of the specialized memory controller present in hardware DSM systems. We then discuss our recent research results in active memory systems showing that our active memory controller design increases single-node performance. These results argue for the inclusion of active memory support in forthcoming commodity workstations. We show that active memory support can be treated as an extension of the cache coherence protocol, and that an active memory controller also contains the necessary functionality for building a hardware DSM machine. Coupled with enhancements in network technology and a small amount of software support, active memory clusters can achieve hardware DSM performance on next-generation commodity servers. 1