search

sakjain92 / Fractional-GPUs

Splits single Nvidia GPU into multiple partitions with complete compute and memory isolation (wrt to performace) between the partitions
149 stars 34 forks source link

Can i using FGPU? #3

Open DonggeunYu opened 2 years ago

DonggeunYu commented 2 years ago

If I apply your repo, can I divide the GPU like this and check it on nvidia-smi? like this: https://docs.run.ai/Researcher/Walkthroughs/walkthrough-fractions/

sakjain92 commented 2 years ago

You will have to try. I worked on this project 3 years ago so I am bit hazy on details but I think it should work. I am not sure if details on nvidia-smi will match but again, you will have to try it out to see. I think the documentation in this repo should be detailed enough to help you carry out the PoC. Please let me know the results if you do carry out the PoC.

DonggeunYu commented 2 years ago

@sakjain92 FGPU evaluation.sh worked successfully, but I couldn't confirm the GPU split with nvidia-smi. As in TODO.md, GPU virtualization seems to be able to split the GPU into containers. But I don't know the direction.

sakjain92 commented 2 years ago

I don't think nvidia-sm will actually show the split. This is because the way I have implemented GPU partitioning is hidden from nvidia-smi.

Does the evaluation.sh script shows you that multiple benchmark applications can run in parallel with isolation? If yes, then look into evaluation.sh script and you can run your applications in similar manner as to how the script is running the applications. This way, your applications should also be able to achieve isolation Note: In my current GPU partitioning applications, you have to modify the application code to make it partition-able (The system is not transparent to applications). So do look at the benchmark application's code also.

DonggeunYu commented 2 years ago

Let's analyze it. Thank you~

sakjain92 commented 2 years ago

Let me know what results do you see. Thanks.