Exploiting Fine-Grain Concurrency Analytical Insights in Superscalar Processor Design

This dissertation develops analytical models to provide insight into various design issues associated with superscalar-type processors, i.e., the processors capable of executing multiple instructions per cycle. A survey of the existing machines and literature has been completed with a proposed classification of various approaches for exploiting fine-grain concurrency. Optimization of a single pipeline is discussed based on an analytical model. The model-predicted performance curves are found to be in close proximity to published results using simulation techniques. A model is also developed for comparing different branch strategies for single-pipeline processors in terms of their effectiveness in reducing branch delay. The additional instruction fetch traffic generated by certain branch strategies is also studied and is shown to be a useful criterion for choosing between equally well performing strategies. Next, processors with multiple pipelines are modelled to study the tradeoffs associated with deeper pipelines versus multiple pipelines. The model developed can reveal the cause of performance bottleneck: insufficient resources to exploit discovered parallelism, insufficient instruction stream parallelism, or insufficient scope of concurrency detection. The cost associated with speculative (i.e., beyond basic block) execution is examined via probability distributions that characterize the inherent parallelism in the instruction stream. The throughput prediction of the analytic model is shown, using a variety of benchmarks, to be close to the measured static throughput of the compiler output, under resource and scope constraints. Further experiments provide misprediction delay estimates for these benchmarks under scope constraints, assuming beyond-basic-block, out-of-order execution and run-time scheduling. These results were derived using traces generated by the Multiflow TRACE SCHEDULING™(*) compacting C and FORTRAN 77 compilers. A simplified extension to the model to include multiprocessors is also proposed. The extended model is used to analyze combined systems, such as superpipelined multiprocessors and superscalar multiprocessors, both with shared memory. It is shown that the number of pipelines (or processors) at which the maximum throughput is obtained is increasingly sensitive to the ratio of memory access time to network access delay, as memory access time increases. Further, as a function of inter-iteration dependency distance, optimum throughput is shown to vary nonlinearly, whereas the corresponding Optimum number of processors varies linearly. The predictions from the analytical model agree with published results based on simulations. (*)TRACE SCHEDULING is a trademark of Multiflow Computer, Inc