Shared-memory Metacomputing

University Of Maryland, College Park, College Park MD

Investigators

Abstract

The focus of this research is on developing techniques and algorithms that will allow idle workstations to be opportunistically exploited for parallel and distributed shared-memory applications. The target environments range from local-area networks (LAN's) to wide-area networks (WAN's). The central challenges are competition from other applications for workstation and network resources, load imbalance arising from differing system speeds, and heterogeneity. The intent is to design and build as transparent a system as possible. Such a system will need new methods of deriving dynamic sharing and processor capacity information online, new load-balancing and thread-migration heuristics that minimize both communication and load imbalance in dynamic environments, and easy-to-use API's that will provide the type information needed to address heterogeneity. The end result of this work will be a set of techniques, policies, and code that will allow shared-memory applications to be transparently migrated across a set of non-dedicated workstations. In allowing otherwise-unused resources to be exploited by this new application domain, this research will significantly expand the value of incrementally-deployed sets of workstations.