The main code to run in order to get the results is in Run-Experiments.ipynb. Here you can configure
the parameters such as number of epochs, place to save the logs, etc...
In order to be able to run the script you need to have accessible your GCLOUD private key (normally
stored in the .ssh folder in linux), and also create a service account and a .json key for the
libcloud driver.
The experiments to run are saved in .csv format in the experiment_designs folder, both for the
2k factorial as for the full factorial.
In case you want to tweak or change those files you can just edit the create-experiments.ipynb
notebook, and change the factors and levels directly there.
To get the measurements from the different servers, you need to copy the cpu_io_stats.py files to
your servers, and specify in the main notebook the location of this file. This outputs a pickle file
with the cpu and io wait percentages every interval seconds.
In the experiment-results notebook we provide code to read the experiment files and perform the ANOVA analysis of the data.
We also provide there the code used to model the system with queuing theory and the final model using multiple random forests.