Advanced Reinforcement Learning for Cloud Resource Scheduling

Overview

This project implements a PPO-based adaptive cloud scheduler that dynamically allocates CPU and memory resources across multiple compute nodes. The environment simulates workload arrivals and cluster-level scheduling decisions.

Features

• Multi-node cloud cluster simulation
• Poisson workload generator
• SLA-aware reward function
• Load balancing penalty
• PPO reinforcement learning agent
• Tensorboard experiment tracking
• Docker reproducibility

Algorithm

Proximal Policy Optimization (PPO)

State Space

• Per-node CPU and memory utilization
• Incoming task CPU and memory demand
• Global queue backlog

Reward Design

The reward function optimizes:

• Resource utilization
• Task latency
• SLA violations
• Load imbalance

Training

Run training:

python train_ppo.py

Evaluation

Run evaluation:

python evaluate.py

Tensorboard

tensorboard --logdir logs

Docker Execution

docker build -t cloud-rl-advanced . docker run cloud-rl-advanced

Conclusion

The trained PPO agent learns scheduling strategies that improve utilization while maintaining low task backlog and balanced cluster load.