https://github.com/aukarande/PACT

Authors

Aditya Ukarande, Dr. Toygun Basaklar, Mingcong Cao, Dr. Umit Ogras

• About PACT
• Quickstart
  • Installation
  • Estimate Carbon Emission
  • Packages Required By PACT
• Workloads
  • Stress Tests
  • Computer Vision Benchmarks
  • Natural Language Processing Benchmarks
  • Reinforcement Learning Benchmarks

About PACT

PACT offers a methodology for accurately analyzing power consumption and tracking carbon emissions of a specific hardware setup. By utilizing PACT, we can make informed decisions aimed at optimizing energy efficiency, minimizing carbon footprints, and advancing toward a greener future.

Quickstart

Installation

Add PACT tracker to your python script (/src/trackerPACT.py):

currentDir = os.path.dirname(os.path.realpath(__file__))
parentDir = os.path.abspath(os.path.join(currentDir, os.pardir))
sys.path.append(os.path.join(parentDir, "src"))
from trackerPACT import PACT

Collect Performance Counters and Power (W)

events_groups = [['MIGRATIONS'],['FAULTS'],['CACHE-MISSES'],['CYCLES']]

@PACT(measure_period = 1, perf_measure_period = 0.01, events_groups = events_groups, tracker_file_name = "./PACT.csv", PSU = "Corsair_1500i")
def your_function():
  # your code

• measure_period: Specify the sampling period (in seconds) to measure power.
• perf_measure_period: Specify the sampling period (in seconds) to collect performance counter data.
• events_groups: Specify the performance events to monitor during execution.
• tracker_file_name: Specify the name of the file to store power and performance data.
• PSU: Specify the power supply unit being used for measurement. Current support for 'Corsair_1500i' and 'NZXT_850'.

Packages Required By PACT

1. liquidctl: https://github.com/liquidctl/liquidctl.git
2. perfmon: https://github.com/VitorRamos/performance_features.git
3. pyNVML: https://pypi.org/project/pynvml/

Workloads

PACT was evaluated by subjecting it to various workloads, including stress tests designed to pressure different hardware components. Additionally, state-of-the-art machine learning models from computer vision, natural language processing, and reinforcement learning were used to assess PACT's performance under real-life workload scenarios.

Stress Tests

We stressed various hardware components using the following tools:

Tools Required:

1. sleep: https://man7.org/linux/man-pages/man3/sleep.3.html

Idle Test

2. stress: https://linux.die.net/man/1/stress

CPU, File IO, Virtual Memory (VM), and HDD Tests

3. gpu-burn https://github.com/wilicc/gpu-burn

GPU Tests

Computer Vision Benchmarks

Source Repository: https://github.com/kuangliu/pytorch-cifar

Applied our tracker to training script in the above repo, and collected emissions and performance metrics while training DenseNet, VGG, ResNet, and MobileNet on CIFAR10 classification task.

Natural Language Processing Benchmarks

Source Repository: https://github.com/aminbigdeli/Text-Classification-using-Transformers

Applied our tracker to training script in the above repo, and collected emissions and performance metrics while fine-tuning bert models on SST-2 classification task.

Reinforcement Learning Benchmarks

Source Repository: https://github.com/denisyarats/pytorch_sac

Applied our tracker to training script in the above repo, and collected emissions and performance metrics while training the agent in cheetah-run, finger-turn-hard, humanoid-run, and quadruped-run environments.