recent results	introduction	E&H resources

Future plans

In the near-term future, we expect to complete a detailed comparison between Cashmere-2L and the Shasta coherence system developed by Kourosh Gharachorloo and Dan Scales of Digital's Western Research Lab. We also expect to complete a version of Cashmere-2L that does not rely on the total ordering or broadcast features of the Memory Channel, allowing us to evaluate the impact of those features on coherence system performance. To accommodate applications (including data mining and object recognition) with very large problem sizes, we will extend the Cashmere system to seamlessly incorporate out-of-core data. Finally, we plan to return to the Cashmere/TreadMarks comparison, evaluating the extent to which one needs to track the strict happens-before relationship on a low-latency system, and re-assessing the tradeoffs between master home node copies and distributed diff logs.

Over the somewhat longer term, we will move our compiler infrastructure from TreadMarks over to Cashmere, and incorporate it in day-to-day use. We will also address a variety of system image issues -- in particular the provision for consistent file access across a distributed cluster. Finally, we plan to extend our protocols to encompass a third level of coherence: processes connected to a Cashmere system over potentially long-haul message-passing networks. A "Cashmere-3L" system would, for example, allow the graphical front-end of a high-performance simulation to access data structures residing on a remote clustered server. We expect to continue our work with both data mining and object recognition applications. We also hope to re-build ties to economics and biology in the wake of Wei Li's departure.

Other issues that we hope to reach in the third year of the grant include fault tolerance, architectural heterogeneity, and hybrid protocols. The first two of these become increasingly important on larger or more geographically-distributed systems. Fortuitously, the inherent data replication of S-DSM provides a natural base for fault tolerance via redundancy [HuSc96]. Heterogeneity becomes a significant challenge if machines do not share the same basic data representations.