Efficiently supporting shared memory on top of commodity hardware is one of the greatest challenges to widespread acceptance of parallel computing. Coherence is the central problem: data in a large-scale system must be replicated and kept near to the processors that use them, but changes must be communicated to all of the copies. This research addresses this problem through a distinctive combination of software and low-cost hardware, and through an active integration of "behavior-driven" coherence protocols with compiler-based program analysis. Applications to be studied include collaborative work with researchers in Economics, Biology, Radiology, Genetics, data mining, computer vision, and virtual reality. The hardware base for the research is a cluster of eight multiprocessors (32 processors total) acquired under a collaborative agreement with Digital Equipment Corporation, and connected by DEC's new Memory Channel network. By directly accessing remote memory, the network allows processors to interact two orders of magnitude faster than they can in comparable systems. Software coherence protocols can exploit this fast communication to maintain directories, implement synchronization, and avoid interrupting the work of remote processors. The ultimate goal of the research is to obtain supercomputer performance for both traditional and emerging applications on networks of commodity machines.
