Design and analysis of a 3-dimensional cluster multicomputer architecture using optical interconnection for petaFLOP computing

摘要

In this dissertation, the design and analyses of an extremely scalable distributedmulticomputer architecture, using optical interconnects, that has the potential todeliver in the order of petaFLOP performance is presented in detail. The designtakes advantage of optical technologies, harnessing the features inherent in optics,to produce a 3D stack that implements efficiently a large, fully connected system ofnodes forming a true 3D architecture. To adopt optics in large-scale multiprocessorcluster systems, efficient routing and scheduling techniques are needed. To thisend, novel self-routing strategies for all-optical packet switched networks and on-linescheduling methods that can result in collision free communication and achieve realtime operation in high-speed multiprocessor systems are proposed. The system is designedto allow failed/faulty nodes to stay in place without appreciable performancedegradation. The approach is to develop a dynamic communication environment thatwill be able to effectively adapt and evolve with a high density of missing units ornodes. A joint CPU/bandwidth controller that maximizes the resource allocation inthis dynamic computing environment is introduced with an objective to optimize thedistributed cluster architecture, preventing performance/system degradation in thepresence of failed/faulty nodes. A thorough analysis, feasibility study and description of the characteristics of a 3-Dimensional multicomputer system capable of achieving100 teraFLOP performance is discussed in detail. Included in this dissertation isthroughput analysis of the routing schemes, using methods from discrete-time queuingsystems and computer simulation results for the different proposed algorithms. Aprototype of the 3D architecture proposed is built and a test bed developed to obtainexperimental results to further prove the feasibility of the design, validate initial assumptions,algorithms, simulations and the optimized distributed resource allocationscheme. Finally, as a prelude to further research, an efficient data routing strategyfor highly scalable distributed mobile multiprocessor networks is introduced.