Pytorch-lightning: Multi-GPU Training GPU Usage

[zhenhuahu]

❓ Multi-GPU Training GPU Usage

Before asking:

1. search the issues.
2. search the docs.

Hi, I'm using lightning and ddp as backend to do multi-gpu training, with Apex amp (amp_level = 'O1'). The gpu number is 8. I noticed that during training, most of time GPU0's utilization is 0%, while others are almost 100%. But their memory usage are the same. Is this normal? I use OpenPAI and have attached their utilization and memeory usage below. Thanks.

Code

What have you tried?

What's your environment?

• OS: [e.g. iOS, Linux, Win]
• Packaging [e.g. pip, conda]
• Version [e.g. 0.5.2.1]

[awaelchli]

check the cpu usage, to make sure dataloading is not a bottleneck.

[zhenhuahu]

check the cpu usage, to make sure dataloading is not a bottleneck.

Hi, thanks for the reply. The total metric(including cpu, gpu) is as follows:

From these images, I have no idea if dataloading is a bottleneck.

I also did profilers, as mentioned in this issue . The dataloading time is constant. But there's some weird call to ''torch._C._TensorBase' objects' that takes a lot of time.

Thanks!

[awaelchli]

Your cpu usage seems high. It could be the cpu is the bottleneck here. Try fewer gpus and observe then observe the gpu utilization.

[zhenhuahu]

Your cpu usage seems high. It could be the cpu is the bottleneck here. Try fewer gpus and observe then observe the gpu utilization.

Yes. I have found it out. Thank you so much!