GaugePy

A highly user-friendly Python library for measuring hardware performance counters on both CPU and GPU.

Prerequisites

• Linux Perf tool
• Nivdia GPU with capability >= 70

Installation

git clone https://github.com/LixiangHan/GaugePy.git
cd GaugePy
pip install -e gaugepy

Usage

Note: This tool requires root/sudo privileges to access hardware performance counters. Please run your scripts with sudo -E command.

CPU HPC Profiling

import gaugepy

def test():
    a = 0
    for i in range(1000):
        a += i

available_cpu_metrics = gaugepy.available_cpu_metrics()

print(available_cpu_metrics)

print(gaugepy.cpu_hpc_profiling(test, available_cpu_metrics))

GPU HPC Profiling

import gaugepy
import torch

def test():
    x = torch.randn(1000, 1000, device="cuda")
    y = torch.randn(1000, 1000, device="cuda")
    z = x + y
    return z

print(gaugepy.gpu_hpc_profiling(test, ["smsp__warps_launched.avg"]))

Citation

If you find GaugePy useful in your research, please consider citing:

@misc{gaugepy,
  author = {Lixiang Han},
  title = {GaugePy},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/LixiangHan/GaugePy}},
}

Disclaimer

This library is not fully tested and may not work on all platforms. If you find any issues, feel free to contribute to this project.