Measurement-Based Performance Models of Numerical Problems on Loosely-Coupled Systems

95 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1988.Multiprocessor systems offer the potential of significant performance improvements over uniprocessors. Performance evaluation is necessary to effectively utilize these systems and to uncover and remove performance bottlenecks, which prevent the realization of their performance potential. Multiprocessors fall into one of two categories: shared and distributed memory systems. We study the performance of distributed memory, or loosely coupled, systems in the context of iterative numerical problems.We derive a performance model for a bus-based system and compare the performance predicted by the model to data collected from experiments implemented on a network of workstations connected via a local area network. We show that the effectiveness of such a system is limited by the bandwidth of the bus.We derive a similar model for the hypercube system and compare it to experimental data collected on the Intel iPSC hypercube. We show that the data downloading and the collection of results reduce the speedup especially for small problem sizes and present schemes to overcome the overheads of data downloading and result collection.Both the bus and the hypercube models assume contention-free and nearest-neighbor communication. We derive a general (architecture and problem independent) model that includes communication contention and allows for non-nearest neighbor interactions. We show that the performance limiting factors are the communication latency, the communication decomposition, the channel congestion, and the grain size.Parallel systems are more vulnerable to faults due to their multiplicity of hardware. Algorithm-based fault tolerance has been proposed for low-cost detection of errors and location of faulty processors in distributed computation. In the last part of the thesis, we study the effectiveness of some of the proposed algorithm-based fault tolerance schemes on the hypercube.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD