• For testbed setup and experiments please refer to TESTBED.md

• This has been tested on Ubuntu 22.04 with Python 3.10.12.
• It is assumed that Python 3.10.12 can be invoked with python3.

• Run bash setup.sh and follow all prompts.

• Run jupyter notebook regenerate_plots.ipynb and execute the cells to (re)generate plots from our paper

• Run python3 simulation/sim.py -workload themis1 -num_gpus 64 -num_apps 500 -scheduling_policy FIFO -output_file FIFO_toy_result.csv to run a simulation with a FIFO scheduler, 500 applications sampled from the themis workload on a 64 GPU cluster
• Run python3 simulation/utils/result_summary.py -fnames FIFO_toy_result.csv to get the summary statistics

• Run bash run_toy_example.sh to run and compare FIFO, SRSF, PCS-pred, PCS-bal, PCS-jct with the same workload/configurations as in the previous example

• Run bash run_workload2.sh
• Once the script finishes, you can use figure7_plot.ipynb to analyze and plot figure 7a,7b

• Run bash run_workload3.sh
• Once the script finishes, you can use figure8_plot.ipynb to analyze and plot figure 8a,8b

• Run bash profile_time_per_sim.sh
• Once the script finishes, you can use figure9a_plot.ipynb to analyze and plot figure 9a

• Run bash profile_sensitivity_error_in_size.sh
• Once the script finishes, you can use figure10_plot.ipynb to analyze and plot figure 10a,10b

• Run python3 reproduce.py -h to see which traces and workloads are acceptable for correct usage.
• It will also show you additional simulation configuration parameters.

The PDF of the paper is also available

• Upon running python3 reproduce.py <workload>, you will be shown an interactive figure showing the pareto front.
• Click a point on the pareto front and click Create Config. This will generate a configuration file for a simulation which will then run. The simulation results are displayed at the end.