Code and dataset for Interference-Aware Edge Runtime Prediction with Conformal Matrix Completion.

1. Setup: Assuming you have a nvidia GPU, you can simply install all dependencies with conda and poetry:

   conda create -n pitot python=3.12
   poetry install

   • The processed dataset used in the paper is included in the data/ folder.

2. Train Models: To replicate the experiments shown in the paper:

   make splits
   make experiments

   • make experiments calls python manage.py train, which will automatically run all experiments from the specified list which are not present in results/. Note that each method result folder will contain a config.json with all hyperparameters used.
   • See pitot/_presets.py for the programmatically generated list of experiments.
   • This assumes that python points to the environment python which poetry install was run inside; if this is not the case, you can also modify the makefile with PYTHON=your/python/bin.

3. Evaluate: The results are evaluated summarized to make them easier to analyze later:

   make evaluate -j16  # or some other number of jobs
   make alias

   • This runs on CPU, so can be run simultaneously while training is in progress.
   • Some runs are shared between multiple ablations, so are alias'd together for convenience with make alias.

4. Analyze: The plots shown in the paper can be generated with

   make figures 

See here for the benchmarks used to collect our dataset.

Dataset conventions:

• t: float[N]: observed average runtime (execution time), in nanoseconds.
• i_{axis}: int[N]: index along each axis into matrix/tensor axes.
• d_{axis}: float[axis:len, axis:features]: side information for each axis.
• n_{axis}: str[axis:len]: names corresponding to values in each axis.
• f_{axis}: str[axis:features]: feature names corresponding to side information for each axis.

Dataset entry names:

• data.npz: workload, platform; contains d_{}, n_{}, f_{} data/metadata.
• if2.npz: workload, platform, interference0
• if3.npz: workload, platform, interference0, interference1
• if4.npz: workload, platform, interference0, interference1, interference2

Python modules:

• prediction: core resuable library runtime prediction functionality.
• pitot: implementation of pitot and other baselines shown in our paper.
• Core python code is type-annotated wherever possible, and can be statically typechecked with mypy.

Scripts:

• preprocess.py / preprocess: data preprocessing steps to turn raw data into .npz dataset files.
• manage.py / scripts: main split, training, and evaluation scripts.
• plot.py / plot: scripts for drawing the figures shown in the paper.

Data files:

• data: our dataset is included in this repository directly.
• splits: this directory is created by make splits.

Result files:

• results: outputs of each experiment are saved here. Has the following structure:

  results/
      method/path/     # can be an arbitrary number of levels
          config.json  # configuration parameters used for training
          0.1/         # data split size
              0.npz    # training log for replicate 0
              0.pkl    # weights for replicate 0
              ...
          ...
      ...

• summary: output of method evaluation.
  • Each results/method/path/ is turned into one summary/method/path.npz.
  • Entries are stacked with data split size as the first axis and the replicate as the second axis (on top of any remaining axes, e.g. target quantile)