Date of Award
Doctor of Philosophy (PhD)
To analyze the performance of applications and architectures, both programmers and architects desire formal methods to explain anomalous behavior. To this end, we present various methods that utilize non-intrusive, performance-monitoring hardware only recently available on microprocessors to provide further explanations of observed behavior. All the methods attempt to characterize and explain the instruction-level parallelism achieved by codes on different architectures. We also present a prototype tool automating the analysis process to exploit the advantages of the empirical and statistical methods proposed. The empirical, statistical and hybrid methods are discussed and explained with case study results provided. The given methods further the wealth of tools available to programmer's and architects for generally understanding the performance of scientific applications. Specifically, the models and tools presented provide new methods for evaluating and categorizing application performance. The empirical memory model serves to quantify the hierarchical memory performance of applications by inferring the incurred latencies of codes after the effect of latency hiding techniques are realized. The instruction-level model and its extensions model on-chip performance analytically giving insight into inherent performance bottlenecks in superscalar architectures. The statistical model and its hybrid extension provide other methods of categorizing codes via their statistical variations. The PTERA performance tool automates the use of performance counters for use by these methods across platforms making the modeling process easier still. These unique methods provide alternatives to performance modeling and categorizing not available previously in an attempt to utilize the inherent modeling capabilities of performance monitors on commodity processors for scientific applications.
Cameron, Kirk William, "Empirical and Statistical Application Modeling Using on -Chip Performance Monitors." (2000). LSU Historical Dissertations and Theses. 7247.