A scheduler log replayer and logs for BFTrainer evaluation
收藏资源简介:
This is a temporary repository to host source for paper: <code>BFTrainer: Low-Cost Training of Neural Networks on Unfillable Supercomputer Nodes.</code> Files <code>BFTrainer-replay.py</code> the main program to replay real scheduler to evaluate our resource allocation algorithm <code>jobs.py</code> implements functions to manage jobs. <code>progCBC.py</code> or <code>progGRB.py</code> the implementation of mixed Integer linear programming using Gurobi optimizer (progGRB.py). We also open source our implementation (progGRB.py) using free optimizer (e.g., CBC, Pulp and JuMP). You can get an Trial Licenses or Free Academic Licenses from Gurobi if you want to run the current version. You need to adjust the import source in the <code>BFTrainer-replay.py</code> to use the CBC based solver. <code>trace.py</code> has functions to manage scheduler logs for the replay evaluation.



