| External factors such as unpredictable behavior of computers and failures complicate the effective use of distributed multiuser environments as a parallel processing platform. This dissertation presents a unified set of techniques to build a metacomputer, that is, a reliable virtual shared-memory computer, on a set of unreliable computers. The techniques are specifically designed for adaptive execution of parallel programs on dynamic and faulty distributed environments. The dissertation presents four software systems to validate the feasibility of these techniques, both for networks of commodity workstations and for the World Wide Web, which lacks a shared file system and user-access control.; Calypso is a programming and a runtime system to address the difficulties of parallel programming on networks of workstations. The parallelism expressed by Calypso programs reflect the problem rather than the execution environment. The Calypso runtime system adapts computations to efficiently use the available resources: the number of workstations may grow and shrink dynamically, and the workstations may fail and slowdown at unpredictable times. ResourceBroker is a resource manager to facilitate the use of otherwise idle workstations. ResourceBroker demonstrates a novel technique to dynamically manage adaptive programs that were not designed to work with external resource managers. As a result, a mix of adaptive programs, written using diverse programming systems can execute side-by-side on a set of workstations.; Charlotte leverages the code-mobility and secure execution of Java applets to extend the concept of metacomputing to the World Wide Web. Charlotte is the first parallel programming system to provide volunteer-based one-click computing: it allows any user on the Internet, without any administrative effort, to participate in ongoing computations by a simple click of the mouse. The Charlotte's runtime system transparently provides load-balancing and fault-masking. KnittingFactory is a software infrastructure that facilitates the execution of Charlotte programs. The contributions of KnittingFactory include a distributed directory service that migrates most of the computations to the client side, and a middleware service for direct applet-to-applet communication. |
