Awesome

GPU relative performance

This is a solution to a problem that if you probably have not even encountered if you are not running a multiple different GPU's in one way of another... of course, this package tries to solve that!

First world problems.

Be it for Neural Network training or Mining you need to have a way to distribute the load equally to each GPU. For example although mxnet supports multiple GPU's the scheduler currently does not know the relative performance of each GPU available and either you have to do it by hand or a uniform load distribution is applied to each GPU (more here).

The (quick) solution.

Now, I've thought of how to tackle this problem with micro-benchmarks, load measurements and so on... but I like the KISS principle and hence I am using something simple with minimal overhead which works surprisingly well in practice. Basically what I want to accomplish is to get a relative performance of each GPU against the lowest performing one currently installed; this can be achieved by using Geekbench aggregated CUDA benchmarks scores and construct a relative performance index for the currently installed GPU's.

The main gist of this solution is fetch the raw JSON CUDA benchmark data from Geekbench, parse it, find the GPU's installed in the system while matching and normalizing their performance using the CUDA benchmark scores. These results can then be immediately used in the mxnet scheduler as percentages.

Requirements

Currently this exists as a source distribution and requires nvidia-smi to be installed -- which if you are using an NVIDIA GPU with either Windows or Linux it should already be installed. Roughly, the requirements are as follows:

* Python > 3
* Nvidia Drivers (in scope)

Unfortunately, MacOS does not have nvidia-smi yet, but a workaround exists which I will probably include in a future update.

Usage

Using this package is easy, first of all do:

$ pip install gpurelperf

Then, after install completes you can use it as a normal package:

from gpurelperf import get_sys_cards

print(get_sys_cards())

A quick example with mxnet

import mxnet
from gpurelperf import get_sys_cards

# returns a tuple with the ratios
(wl_list, gfx_list) = get_sys_cards()

# then you would set the work load list as such
work_load_list=wl_list

License

This project is licensed under the terms and conditions of the Apache 2.0 license.