Michael L. Scott, PI (scott@cs.rochester.edu)
Sandhya Dwarkadas, co-PI (sandhya@cs.rochester.edu)
Department of Computer Science
University of Rochester
Rochester, NY 14627-0226
The Computer Science Department at the University of Rochester has a long history of research in shared-memory parallel computing. Much of this research has been motivated by the desired to maximize price-performance for "knee-of-the-curve" parallel systems -- to obtain the best possible performance consistent with modest hardware costs. Through the early 1990s a series of simulation-based studies [KoS95a, KoS95b, KoS95c, KoSB95, KoSc96] suggested that the coherence problem -- the need to keep cached copies of data consistent in the face of remote changes -- could be solved efficiently and economically in software, provided that the underlying hardware provided non-coherent access to the memory of remote nodes.
The funded research aims to demonstrate this hypothesis in practice through experimental evaluation of a working coherence system on a cluster of small-scale multiprocessors connected by a network -- the Digital Memory Channel -- that provides low-latency non-coherent writes, from user space, to the memory of remote nodes. Our current hardware base, provided through a combination of NSF RI funding and a half-million dollar equipment grant from Digital, consists of eight 4-processor AlphaServer nodes. Our software base consists of the locally-developed Cashmere coherence system, locally-modified variants of the TreadMarks coherence system, and a variety of compiler tools.