> question about multiple GPU with mpnn

when i put mpnn.py in the two GPU case, it seems not work, batch data issue, and data split problem, so can u confirm it do not work? thanks.

Posted by: EversChen @ Aug. 12, 2019, 11:08 a.m.

Hi. Our current Alchemy toolkit (both PyG and DGL) does not support mult-gpu. We welcome contributions of a multi-gpu implementation.

Posted by: jiezhongqiu @ Aug. 13, 2019, 6:42 a.m.

thanks for feedback

Posted by: EversChen @ Aug. 13, 2019, 10:02 a.m.

We installed two GPUs for one PC. I tried to use two GPUs to speed up the training of a model, but after testing, it was found that the speed was not significantly improved compared to the single GPU. I guess it's because the dual-GPU interconnect configuration was not carefully optimized, so I didn't continue to spend more time on this in this competition.
Considering that different models need to be trained to predict different properties, we chose to have different GPUs separately trained to achieve parallel acceleration. Compared to using 2 GPUs to train one model, this can make full use of the performance of the single GPU.

Posted by: Masker_L @ Aug. 13, 2019, 1:10 p.m.

so you need two PC to make use of two GPU as you said, right? i understand your solution.
however i don't think two GPU can't improve the train, are you sure your two GPU on one PC is working?
and what's your model? you know for parallel handling, you have to update your model to adapt it.

Posted by: EversChen @ Aug. 13, 2019, 1:59 p.m.

2 GPU with one PC.
I have tried multi-GPU methods in other tasks to speed up, even if the model is adapted accordingly, the accelerated performance is limited.
I don't mean 2 GPU cat't improve the performance. Even if you can make the two GPUs work together perfectly, it will reach 160~170% of the performance of the single card at most, which requires you to configure the hardware.
In this competition, I think it is better to use two GPU to run separate optimization tasks for different properties. Just for advice.
Really admire you for the geek spirit, good luck!

Posted by: Masker_L @ Aug. 13, 2019, 2:53 p.m.

U are right, for this specific task we can just train two separated models separately, since we never know how many issues there on the way, we have to save time, as the remain time is limited, thanks for hint.

Posted by: EversChen @ Aug. 13, 2019, 3:14 p.m.
Post in this thread