Content area
Network distributed computing has become widely used for scientific and engineering computing. With the advances of high-speed general-purpose network and communication technology, geographically distributed computing has become possible and the desire for it is more and more apparent. The critical element in network distributed computing is the communication network, and its performance is dependent upon the performance of communication.
In order to improve the communication performance, it is important to understand the network traffic characteristics generated by distributed applications. This dissertation presents the results from a CFD code ENG10. The characteristics include bit rate, burstiness, interarrival time, and burst duration. Moreover, the communication performance is also affected by system software, API and algorithms used for distributed applications. Those effects can be understood by evaluating time distributions during executing distributed applications. The time distributions presented in this dissertation include MPI setup time, communication time, calculation time, their ratios to the execution time and the speedup for a variety of the number of nodes and the size of messages over both ATM and Fast Ethernet networks.
In geographically distributed computing, the signal propagation delay becomes one of the major factors that affect the communication performance. This delay is difficult to reduce by using faster networks. This dissertation designs and develops a Shared Object Space (SOS) to improve the communication performance in geographically distributed computing by providing protocols to support prefetching. By using this method, SOS is able to make remote objects available locally, thus reduce the communication delay. SOS has an object-oriented, shared memory programming model interface. The performance of SOS has also been shown and discussed in this dissertation.