Multiple channel architecture: A new optical interconnection strategy for massively parallel computers
Abstract
A major problem in designing large-scale parallel systems is the construction of an interconnection network to provide interprocessor communication and memory access for the many processors involved. Current device technologies and interconnection topologies that have been implemented or proposed for large scale multiprocessors are described in detail, listing each proposal's strengths and weaknesses. A parallel processing architecture that allows multiple interconnection topologies using optical frequency division multiplexing is proposed to provide the communication bandwidth required for large multiprocessor computer systems. Data interchange connections between devices is established by assignment to the same channel. The existence of many channels of communication and the ability to quickly switch between them allows hierarchies of architectures and the simultaneous execution of multiple parallel tasks. Dynamic assignment of arbitrary numbers of processors, memories, I/O devices, and communication channels allows many new forms of parallel computation. A detailed simulation software system that was developed to study the proposal is described. Using this simulator, detailed analysis of cache performance, processor execution, memory access, and virtual bus are possible. Three parallel programs (matrix multiply, Gauss elimination, and median filtering) are studied on the simulator. Results indicate that non uniform shared memory accesses significantly reduce the utilization of individual processors. However, unlike multistage interconnection networks, data traffic not involved with memory hot spots is not affected. This architecture's performance is found to be sensitive to the CSMA/CD backoff parameter, and transmission speed of the medium. Performance is not significantly affected by bus loadings up to 8 nodes per bus.
Degree
Ph.D.
Advisors
Meyer, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.