| Increasingly powerful workstations and PCs, interconnected through various networks, continue to proliferate throughout the world. Most are greatly underutilized and thus represent a significant computational resource, which could be tapped for running applications requiring large amounts of computations. However, developing distributed applications is significantly more difficult than developing sequential applications. Therefore, distributed computing has not become a main stream of computing.; This dissertation presents a distributed computing environment, PODC, on top of a mobile agent system for distributed computing that uses the concept of well-known paradigms. These paradigms provide the communication infrastructure, while the application programmer only supplies application-specific functions through a graphic interface. The PODC system transparently handles the selection of computers, the mapping of workload, and the monitoring of the execution. The five paradigms presently supported in PODC are the bag-of-tasks, branch-and-bound search, genetic programming, finite difference, and individual-based simulation. In all cases, significant speed-ups can be achieved while hiding the details of distribution and parallelism from the application programmer.; The main features of PODC, which differentiate it from other paradigm-based approaches, are the following: (1) It is intended for loosely-coupled network environments, not specialized multiprocessors; (2) it is based on an infrastructure of mobile agents; (3) it supports programming in C, rather than a functional or special-purpose language, and (4) it provides an interactive graphics interface through which programs are constructed, invoked, and monitored.; In addition, this dissertation has also addressed common optimization techniques for each paradigm. In particular, three issues related to iterative grid-based applications, typified by the finite difference paradigm and the individual-based simulation paradigm, have been addressed. The first issue concerns program performance. I introduce a technique called SuperBoundary Exchange for significantly reducing communication overhead by trading redundant computation and larger message size for less frequent communications. The second issue is programmability. I introduce a Global-Indexing Distributed Memory, which allows distributed programs to share a common virtual global user grid. The third issue is the correctness and repeatability of distributed programs. I introduce four random number generation schemes to allow distribution and replication of computation at different granularity without loss of repeatability. |
