This is a toolbox for simulating the heat loop in a small datacenter and providing standard bindings for an RL environment to it. It has two backends, first a very simple one described here and then a more complex CFD based one described here.
Clone this repo to you local drive and install the packages defined in requirements.txt using any desired method, for example
pip install -r requirements.txt
The exact versions used can be seen in versions.txt and has been tested using python 3.6.9 in a linux environment.
First you start the ray daemon depending on your cluster. For a local one-node cluster this is done with
ray start --head --num-cpus=X --num-gpus=Y
where it can be nice to not use all cpus for ray so that tensorboard can work in the background.
For a cluster with multiple nodes there are resources on the ray website.
Finally the simulation is run with
python3 run.py <args>where there are some optional parameters available among which we have
--help # Show all available commands
--rafsine # Use rafsine instead of simple simulation, needs additional git repo with rafsine code
--n_samples NS # How many samples the hyperparameter search should run over
--tag SOME_NAME # Name of the run
--timesteps NT # How many seconds (in simulation) should each sample run
Tensorboard is started with
tensorboard --logdir /path/to/repo/results --reload_multifile true
and can be accessed with a browser at localhost:6006.
Ray dashboard is running with the ray deamon and can be access with a browser at localhost:8265.