| This thesis investigates the design and implementation of a Java-based system to enable Internet Computing (i-Computing ), which is a new type of distributed computing involving heterogeneous and non-dedicated workstations from different organizations on the Internet. Promises of i-Computing include inexpensive supercomputing, large-scale resource sharing, and more efficient use of existing resources.; This thesis first presents the potentials and challenges for i-Computing, and identifies the issues to be addressed. Then an i-Computing system called Distributed Java Machine (DJM) is proposed. DJM is a global distributed virtual machine. The following technologies are applied in building DJM: Internet infrastructure, World Wide Web, distributed computing, Java programming and object-orientation. DJM explores the coarse-grained object parallelism based on under-utilized workstations on the Internet.; As the effective power of machines for i-Computing is heterogeneous and volatile, a novel stochastic load balancing (Stoc) and a job migration algorithms are introduced to improve the efficiency. Stoc combines the transfer and the location policies using a probabilistic function, where the probability of a machine being selected is inversely proportional to its estimated cost. Moreover, owing to reasons such as machine crashes or owner reclamation, i-Computing systems are more vulnerable to failures. Hence, an object-based fault tolerance scheme is proposed. This scheme uses the classifications of messages, methods, objects, and machines to improve its effectiveness and efficiency.; Correctness of the fault tolerance scheme has been proved. The performance of DJM has been evaluated by simulations and empirical experiments. Finally, a general optimization solver named APGAIN, which is based on adaptive parallel genetic algorithm model, has been developed to demonstrate the practical usage and the efficiency of the overall DJM system.; Results obtained from empirical experiments are encouraging. For example, DJM has achieved a speedup of 5 to 8 times by 14 workstations in a LAN environment, and a speed up of 4.5 by 8 workstations in a WAN environment. In the load balancing experiment, Stoc has obtained a speed improvement of 15% to 20% over other well-known load balancing approaches. Using the job migration facility of DJM, an application has achieved a speed improvement of 32% over the non-migratory counterpart. Moreover, the scalability test has shown that DJM can give an application a speed-up curve that is close to linear.; This thesis has made four major contributions in the design and implementation of the following: (a) a general-purpose Java-based platform (DJM) for Internet Computing, (b) a stochastic load balancing algorithm (Stoc), (c) an object-based fault tolerance scheme, and (d) a self-adaptive parallel GA system on the Internet (APGAIN). |
