cache-policy-evaluation

Evaluate cache replacement and branch prediction policies with gem5 simulations.

This is a tool to simulate different types of workloads on a simple CPU, running a given benchmark on a matrix (Cartesian product) of policies and L1-i/L1-d cache sizes. The output is a simple CSV file that can be later plotted and analyzed; a Jupyter notebook is given.

The CPU is described in simulate.py: it is a simplified model of configs/learning_gem5/part1/simple.py that uses the TimingSimpleCPU model from gem5 with a simple L1-DDR3 memory hierarchy.

For detailed information refer to the docs/ folder:

• It is recommended to start with gem5_on_alpha.md to set up the simulation environment.
• Then, read benchmarks.md to understand how to compile benchmarks and which ones are of interest to you.
• Then, read cache_experiments.md to run your first experiment and learn how to process the data.
• branch_prediction_experiments.md will then explain how to use different branch predictors.
• single_simulations.md and existing_cpus.md are for one-shot simulations.
• If you want to hack on this project, tinkering.md has a few notes to get you up and running on the internals.

Finally, in notebooks/graphs.html you can find a simple analysis of simulation data comparing different cache replacement policies in several scenarios.