GPU Power Prediction

This repository provides the implementation and resources used in the research paper:
"Estimating Power Consumption of GPU Application Using Machine Learning Tool"
by Gargi Alavani Prabhu, Tanish Desai, Sharvil Potdar, Nayan Gogari, Snehanshu Saha and Santonu Sarkar Cite as:
Gargi Alavani Prabhu, Tanish Desai, Sharvil Potdar, Nayan Gogari, Snehanshu Saha and Santonu Sarkar, "Estimating Power Consumption of GPU Application Using Machine Learning Tool", 2024. [Google Scholar]

Abstract

Accurately predicting power consumption of GPU kernels is crucial for optimizing performance and energy efficiency in high-performance computing. This project presents a benchmark suite and a PowerAPI tool for profiling and predicting the power usage of GPU kernels based on machine learning techniques. The suite includes a diverse set of CUDA kernels and datasets, enabling robust training and evaluation of power models. PowerAPI provides an easy-to-use interface for instrumenting CUDA applications and collecting power/energy data, facilitating automated and reproducible measurements for research and development.

Repository Structure

• PowerAPI/: Source code and interface for GPU power measurement.
• Benchmark Suite/: Collection of sample CUDA kernels and benchmarks.
  • Each sample folder contains a README with build/run instructions and key concepts.
• Datasets/: Example datasets for training and evaluating power prediction models.
• Supplementary File.pdf: Additional details and supporting material.

Tools Used

• CUDA Toolkit 11.0+: For compiling and running GPU benchmarks.
• PowerAPI: Custom API for measuring GPU power and energy.
• C++/C/CUDA/SWIG/Makefile/Python/Shell: Multi-language implementation for flexibility and extensibility.

How to Use PowerAPI

1. Add PowerAPI to the library path

   LD_LIBRARY_PATH="/path/of/Folder/PowerAPI"

2. Import PowerAPI in your CUDA file
   At the start of your .cu file, add:

   #include "GPUDevice.h"

3. Instrument your kernel code
   Create a GPUDevice object and wrap your kernel calls:

   GPUDevice g1 = GPUDevice(<GPU Device ID>,<Kernel Name>,<Grid Size>,<Block Size>);
   g1.startReading();
   // <<Cuda Kernel Calls>>
   g1.stopReading();

4. Run your code
   Output is saved in a text file named after the kernel.

   • Power in Watts
   • Energy in milliJoules
     Sample output:

   KernelName,GridSize,BlockSize,MaxPower,MinPower,AvgPower,Time,Energy,
   reluKernel, 86436, 256, 84, 64, 76 , 67603944.000000, 5154,
   

Benchmark Suite Usage

• Windows: Use provided Visual Studio solution files (*_vs<version>.sln).
• Linux: Use makefiles in each sample directory:

  cd <sample_dir>
  make

  Options:
  • TARGET_ARCH=<arch>: Target specific CPU architecture
  • dbg=1: Build with debug symbols
  • SMS="A B ...": Build for specific SM architectures
  • HOST_COMPILER=<host_compiler>: Use a custom host compiler

See individual sample READMEs in Benchmark Suite/ for details and supported architectures.

References

• CUDA Toolkit
• CUDA Programming Guide
• Original Research Article (Google Scholar)

---

For questions or contributions, please refer to the repository issues or contact the authors via their GitHub profiles.