Thermal-Aware Parallel Experiments

This project provides a Python-based framework for benchmarking CPU performance under thermal constraints, with a focus on understanding how parallel workload scaling interacts with temperature, throttling, and computation throughput.

Features

• Parallel CPU workload generation using configurable number of processes
• Real-time CPU temperature monitoring using OpenHardwareMonitor
• Logging of temperature and workload iterations over time
• Automatic force-stop if CPU temperature remains high for a critical duration
• Optional pause-on-overheat mechanism (disabled by default)
• Jupyter notebook for cross-experiment analysis and visualization

Requirements

• Python 3.x
• OpenHardwareMonitor (v0.9.6 recommended)
  Must be located at:
  openhardwaremonitor-v0.9.6/OpenHardwareMonitor/OpenHardwareMonitor.exe
• Python packages:
  • numpy
  • pandas
  • matplotlib
  • psutil
  • requests

Install dependencies with:

pip install numpy pandas matplotlib psutil requests

Usage

1. Run Experiments

To execute tests across different numbers of worker processes:

python experiment_runner.py

This will:

• Run tests with process counts such as 1, 2, 4, ..., 16
• Launch worker processes that perform CPU-intensive matrix multiplications
• Monitor and log CPU temperature every 0.1 seconds
• Wait 60 seconds between each test for cooldown

Each run creates logs in folders like logs_nproc_04/ with a CSV file for each worker process.

2. Analyze Results

Open the following Jupyter notebook:

compare_experiments.ipynb

You can compare:

• Computation count per process count
• Temperature profiles
• Throttling events
• Process scaling efficiency

Temperature Safety

To protect hardware:

• If temperature exceeds a threshold (default: 85°C) for more than 30 seconds, all processes are terminated.
• This adds a software-level layer of thermal protection beyond hardware throttling or shutdown mechanisms.
• An optional pause-on-overheat mode is included but disabled by default. It allows the program to wait when overheating instead of stopping entirely.

Customize the temperature threshold using:

TEMP_THRESHOLD=90.0 python experiment_runner.py

Log Format

Each process generates log_{proc_id}.csv with:

• timestamp – ISO 8601 format
• temperature – Measured CPU temperature (°C)
• throttling – True if the program paused due to high temp, otherwise False
• count – Matrix multiplication iteration count

Note: throttling here refers to application-level pauses, not hardware-level frequency scaling.

Performance Insights

Optimal Number of Processes

The chart below shows total computation count per process count. It was recorded on an Intel(R) Core(TM) i7-1065G7 (4 cores, 8 threads).

• The peak occurs at 4 processes — matching the number of physical CPU cores.
• Performance decreases or plateaus beyond this due to scheduling overhead or thermal limitations.
• The 2-process dip may stem from transient throttling or uneven core allocation.

Thermal Behavior Example

Below is a temperature log for 1 process (logs_nproc_01):

• Initially, the temperature rapidly rises past the 85°C threshold.
• After a short period, the temperature stabilizes around 80°C.
• This stabilization suggests hardware-level thermal throttling is actively regulating the system by reducing CPU frequency or redistributing power.

License

MIT License