Cache-Aware Virtual Page Management

Abstract

With contemporary research focusing its attention primarily on benchmark-driven performance evaluation, studying fundamental memory characteristics has gone by the way-side. This thesis presents a systematic study of the expected performance characteristics for contemporary multi-core CPUs. These characteristics are the primary influence on benchmarking variability and need to be quantified if more accurate benchmark results are desired.

With the aid of a new, highly customizable, micro-benchmark suite, these CPU-specific attributes are evaluated and contrasted. The benchmark tool provides the framework for accurately measuring instruction throughput and integrates hardware performance counters to gain insight into machine-level caching performance. Additionally, the Linux operating system's impact on cache utilization is evaluated. With careful virtual memory management, cache-misses may be reduced, significantly contributing to benchmark result stability. Finally, a popular cache performance model, stack distance profile, is evaluated with respect to contemporary CPU architectures. While particularly popular in multi-core contention-aware scheduling projects, modern incarnations of the model fail to account for trends in CPU cache hardware, leading to measurable degrees of inaccuracy.