Do Modern LLC Replacement Policies Generalize to Modern Workloads?
Research evaluating modern cache replacement policies in ChampSim
Modern cache replacement policies such as Hawkeye and Mockingjay have demonstrated strong results on traditional benchmark suites, but their performance on newer workloads is less understood. This project evaluates six replacement policies across SPEC CPU2006, SPEC CPU2017, and Google Workload Traces (GWT) using a custom ChampSim evaluation framework, automated benchmarking pipeline, and analysis tooling.
Research at a Glance
| Question | Do modern LLC replacement policies continue to outperform LRU on contemporary workloads? |
| Benchmarks | SPEC CPU2006, SPEC CPU2017, GWT |
| Simulator | ChampSim |
| Policies Evaluated | LRU, Hawkeye, Mockingjay, SHiP, SRRIP, DRRIP |
My Contributions
- Built an automated benchmarking pipeline for large-scale ChampSim experiments
- Ported Hawkeye and Mockingjay to ChampSim’s modern class-based API
- Instrumented Hawkeye to measure predictor false-positive rates
- Developed Python scripts for parsing simulation results and generating publication-quality figures
- Co-authored the final paper and analysis with Luciano Galvani
Key Results
We find that the traditional policies which outperform LRU on the classic SPEC benchmark suites, and their variants, fail to generalize to the Google Workload Traces, where the same policies degrade IPC by up to 14% at a 2MiB LLC and converge toward LRU performance as LLC size grows. These results show that the SPEC suite is a poor indicator for eval uating replacement policies in the modern server and cloud computing contexts and suggest that new cache replacement paradigms may be necessary to adapt to the contemporary environment.
Highlights
-
Used PC entropy as a workload characterization metric, showing that modern data center workloads exhibit significantly higher PC diversity than SPEC benchmarks, reducing the effectiveness of PC-based replacement predictors.
-
Demonstrated that replacement policies which consistently outperform LRU on SPEC CPU often fail to generalize to modern Google Workload Traces, with most policies degrading performance on real-world server workloads.
Selected Figures
Overall Performance
Normalized geomean IPC on GWT suite showing degrading performance from all policies but Hawkeye.
Suite Characterization Using an Instrumented LRU
Workload suite characterization including the PC entropy metric
Paper
This page provides a high-level overview of the project. The full methodology, experimental setup, statistical analysis, and discussion are available in the accompanying paper.
→ Read the full paper