Application Specific Optimization of Ethernet Cluster Communication

In this diploma thesis, the implementation of a networking protocol, which is adopted from a previous work, is refined to a state where it is ready for production use. The protocol is tailored for the specific needs in a cluster environment which uses standard ethernet as the interconnection hardware. Thereby focuses is on developing and implementing a flow control which is very lightweigt in terms of processing overhead and copes the challenges of such a tightly coupled network gracefully. Second, the Log(G)P models for parallel computation are examined and it is shown that they do not give a usable definition of the overhead or latency parameters for protocols whose overhead is O(message size)–instead, the P-LogP model is used to characterize the protocol and for comparisons to the TCP/IP protocol suite. Last, it is demonstrated that the protocol gives usable results with real-world scientific applications