CARL SCHEDULER

Credits

• Developers: The CARL Scheduler code was developed by Prathamesh ([email protected]) and Saswat from the Cloud Systems Lab (CSL) at IISc.

• Please cite the following publication if you use CARL:

Prathamesh Saraf Vinayak; Saswat Subhajyoti Mallick; Lakshmi Jagarlamudi; Anirban Chakraborty; Yogesh Simmhan; CARL: Cost-Optimized Online Container Placement on VMs Using Adversarial Reinforcement Learning in IEEE Transactions on Cloud Computing, vol. 13, no. 1, pp. 321-335, Jan.-March 2025, doi: 10.1109/TCC.2025.3528446.

• Code provided in this repository is a research prototype and provided as-is.

Installation

pip install -r requirements.txt

Note:

• For vpsolver, you need to install gurobi and add it to your path.
• For RL training, you need to install gym. Currently, gym has been upgraded to gymnasium, better to use that for speed and efficiency. Ideally port the code to gymnasium.
• The way RL envs are registered and uses is different in the new gymnasium.
• All the libraries such as sympy and all are upgraded, the current code is using the older versions. Best to port them to access the new features and stability.
• I had to hack the library code for gym and RL to work.
• Please try to port the code to latest sympy, gymnasium, imitation, stable-baselines3, etc.

Links:

• Gym: https://www.gymlibrary.dev/index.html
• Gymnasium: https://gymnasium.farama.org/
• A2C/PPO: https://stable-baselines3.readthedocs.io/en/master/index.html
• GAIL: https://imitation.readthedocs.io/en/latest/

Usage

• Refer Makefile for evaluation
• Simulator code is in Simulator/
• RL training code is in RL/
• You need to generate expert trajectories first, refer to generate_expert_traj_vp.py in RL/, need access to Gurobi
• To train RL models, refer to commands.txt in RL/, installed libary versions require changes to the library code. Better to install newer libraries and make them work.