DiffTune is a system for learning the parameters of x86 basic block CPU simulators from coarse-grained end-to-end measurements. Given a simulator, DiffTune learns its parameters by first replacing the original simulator with a differentiable surrogate, another function that approximates the original function; by making the surrogate differentiable, DiffTune is then able to apply gradient-based optimization techniques even when the original function is non-differentiable, such as is the case with CPU simulators. With this differentiable surrogate, DiffTune then applies gradient-based optimization to produce values of the simulator's parameters that minimize the simulator's error on a dataset of ground truth end-to-end performance measurements. Finally, the learned parameters are plugged back into the original simulator.

The full paper is available at: https://arxiv.org/abs/2010.04017

To cite the paper, the BibTeX is:

@inproceedings{renda2020difftune,
    title={DiffTune: Optimizing CPU Simulator Parameters with Learned Differentiable Surrogates},
    author={Renda, Alex and Chen, Yishen and Mendis, Charith and Carbin, Michael},
    booktitle={IEEE/ACM International Symposium on Microarchitecture},
    year={2020},
}

A small-scale demo of DiffTune on a synthetic dataset is available at https://colab.research.google.com/drive/1N5pqyYKxmkuwBIHzTcp8DqqVOCEnqCV4?playground=true

• nvidia-docker

Inside of the source directory, run:

./build.sh

Then, to download the preprocessed BHive dataset and the released artifact, run:

./download.sh

To launch the docker container, inside of the source directory, run:

./run_docker.sh

Once inside the docker container (which you can exit with Ctrl-d or exit), run:

cd difftune
name=experiment-name

to go to the DiffTune code directory and set the experiment name (an arbitrary value that keeps data from different runs separated).

To read the original dataset from BHive and the original parameters from llvm-mca, run the following:

python -m difftune.runner --name ${name} --task blocks
python -m difftune.runner --name ${name} --task default_params

These commands create pickle files in data/${name} with the basic block dataset and default parameter tables respectively.

Next, to generate the simulated dataset, run:

python -m difftune.runner --name ${name} --task sample_timings --sim mca --arch haswell --n-forks 100

This command samples parameter tables and runs them through llvm-mca on Haswell, writing the seed and result to data/${name}/mca-haswell.csv. --n-forks=100 specifies to run 100 sampling workers in parallel, which should be tuned based on compute availability. This command does not ever terminate; when sufficient samples have been collected, just kill it with Ctrl-c, or manually truncate it to the desired length.

With the simulated dataset collected, to train the surrogate, run:

python -m difftune.runner --name ${name} --task approximation  --sim mca --arch haswell --model-name surrogate --device cuda:0 --epochs 6

Then to train the parameter table, run:

python -m difftune.runner --name ${name} --task parameters --sim mca --arch haswell --model-name surrogate --device cuda:0 --opt-alpha 0.05 --epochs 1

Finally, to extract the parameter table to a file (data/${name}/surrogate-model-params-extracted) and evaluate its test error / correlations, run:

python -m difftune.runner --name ${name} --task extract --sim mca --arch haswell --model-name surrogate
python -m difftune.runner --name ${name} --task validate --sim mca --arch haswell --model-name surrogate