Performance analysis on a CC-NUMA prototype
IBM Journal of Research and Development - Special issue: performance analysis and its impact on design
| Hi-index | 0.00 |
The quantitative evaluation of pipeline performance and the quantitative analysis of different design alternatives and related cost/performance tradeoffs are a fundamental aspect of high-performance computer system design. This performance evaluation process requires accurate models of both the pipeline organization and the characteristics of the workload being executed. In this paper we derive general mathematical models and analyses of workload behavior and pipeline performance that provide measures as accurate as detailed trace-driven simulations with the computational efficiency of analytic methods. Our approach is based in part on stochastic marked point processes that completely capture the workload behavior affecting pipeline performance and yield accurate distributions for characterizing the instruction stream contents of different pipeline organizations. These distributions are then used to parameterize a model of pipeline performance that captures the dynamics of instruction execution in the pipeline. Our results show that even simple instances of our general models provide performance measures in excellent agreement with those from a detailed, cycle-based simulation for most of the benchmarks considered. We then use these models to study pipeline performance for different workloads and pipeline organizations. More complex instances of our models are also discussed.