Skip to content

Latest commit

 

History

History
72 lines (47 loc) · 2.52 KB

readme.md

File metadata and controls

72 lines (47 loc) · 2.52 KB
Raw

GPU-utils

This is a small library to use in services, that require information about GPU. It can select GPUs based on available memory in GPU-intensive services.

There are several methods to acquire GPU information. We could use, for example:

  • nvml
  • nvidia-smi
  • torch
  • etc.

This library was developed with extensibility in mind, if we were to change primary method of obtaining GPU information

Quickstart

Installation

All GPUInfoObtainer implementations provides optional dependencies for it to work.

pip install gpu-utilities[torch] # for TorchGPUInfoObtainer
pip install gpu-utilities[nvml] # for NvmlGPUInfoObtainer

Usage example

from gpu_utilities.obtainers.nvml import NvmlGPUInfoObtainer
from gpu_utils.gpu_selector import GPUSelectorFactory, GPUSelectStrategy
import torch

if __name__ == "__main__":
    gpu_info_obtainer = NvmlGPUInfoObtainer()  # Use NVML and nvidia-smi as GPU information provider
    gpus_available = gpu_info_obtainer.gpus_available()  # Check if information about GPU is available

    if not gpus_available:
        raise Exception('GPUs are unavailable')

    selector = GPUSelectorFactory().selector(GPUSelectStrategy.BEST_FIT)  # Create Selector for strategy BEST_FIT

    gpu = selector.select(gpu_info_obtainer.gpus(), 800 * 1024 * 1024)  # Sekect GPU that is able to allocate 800 MB according to strategy
    print(f"gpu {gpu} is the best option now!")

    torch.cuda.device(gpu.to_torch_device())  # Example illustrating usage in torch

GPU info obtaining

There are 2 implemented methods at the moment to obtain information about GPU.

  • TorchGPUInfoObtainer
    • via PyTorch library
    • Deprecated since 0.2.0, since it allocates memory on GPU to be used
  • NvmlGPUInfoObtainer
    • Requires NVML to be installed on your machine
    • Might additionally require nvidia-smi

Selection Strategies

Three very basic strategies to choose GPU implemented. All strategies filter GPU's that can allocate desirable memory span.

  • BEST_FIT (Choose the most loaded GPU (memory-wise))
  • WORST_FIT (Choose the least loaded GPU (memory-wise))
  • RANDOM

They are all accessible via GPUSelectorFactory. Enum GPUSelectStrategy contains all implementations.

Naming of the strategies can be seen as a bit odd. It was inspired by page allocation strategies in the OS, where "BEST_FIT" represented the most effecient use, hence the name.