DeepScale: Predictive DVFS for Virtual Machine Workloads

DeepScale is a predictive power management system that uses deep reinforcement learning to dynamically set per-core CPU frequencies based on system-wide VM activity. Unlike traditional DVFS strategies that react to individual core utilization, DeepScale proactively learns workload patterns across VMs to minimize energy usage while maintaining performance.

---

🌟 Key Features

• ⚡ Predictive DVFS: Learns when workload phases change and adjusts frequency ahead of time.
• 🧠 Deep Q-Learning: Uses system-wide CPU metrics to choose energy-optimal frequencies.
• 📉 Energy-Aware: Incorporates RAPL energy readings into its reward function.
• 🔁 Online Training Loop: Learns continuously during benchmarks via reinforcement signals.
• 📊 Trace Logging: Exports a .jsonl log of system metrics and actions for later analysis.

---

📦 Project Structure

• master.py - Main experiment driver (setup, run, destroy)
• vm_manager.py - Starts, pins, provisions, and benchmarks VMs
• host_metrics_daemon.py - Logs CPU stats and RAPL energy
• cpu_dqn.py - Deep Q-Network for learning and frequency control
• cloud-init/ - VM bootstrap configs
• vm_benchmark_scripts/ - Redis benchmark startup scripts
• logs/ - Output logs and metric traces

---

---

🚀 How to Use

1. Setup Phase (Creates and provisions VMs):

sudo python3 -u master.py setup | tee logs/master.log

2. Run Phase (Starts benchmark, collects metrics, performs DQN control):

sudo python3 -u master.py run | tee logs/master.log

3. Cleanup Phase (Stops and removes VMs):

sudo python3 -u master.py destroy | tee logs/master.log

📈 Output Files

• metrics.jsonl: Per-core metrics + energy per second
• logs/master.log: VM creation, benchmarking, and DQN actions
• trained_model.h5: Checkpointed model state from DQN

📚 Dependencies

Make sure you have the following installed:

sudo apt install qemu-kvm libvirt-daemon-system
pip install tensorflow numpy psutil paramiko scp

🧪 Default Workload

By default, PhaseScale provisions each VM with a YCSB benchmark running on Redis. You can configure this in vm_manager.py or by modifying the cloud-init configs.