Pytorch DataParallel two gpu Hang

maskjp · September 23, 2020, 8:14pm

Hi, I have two lambda dual with two TITAN RTX. One pc works fine. But another pc freeze when using Dataparallel.
I got the not working one last mouth.

I tried the solution in this

But got an error message.

And I also try to disable iommu following the following methods. But I found that the iommu is already disabled by default.

github.com/pytorch/pytorch

Multi-GPU K80s

opened 09:52PM - 23 May 17 UTC

closed 10:27AM - 22 Oct 17 UTC

davidmascharka

I'm having trouble getting multi-gpu via `DataParallel` across two Tesla K80 GPU…s. The code I'm using is a modification of the MNIST example: ``` import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torchvision import datasets, transforms from torch.autograd import Variable from data_parallel import DataParallel train_loader = torch.utils.data.DataLoader( datasets.MNIST('../data', train=True, download=True, transform=transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)) ])), batch_size=256, shuffle=True, num_workers=2, pin_memory=True) class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.fc1 = nn.Linear(320, 50) self.fc2 = nn.Linear(50, 10) def forward(self, x): x = F.relu(F.max_pool2d(self.conv1(x), 2)) x = F.relu(F.max_pool2d(self.conv2(x), 2)) x = x.view(-1, 320) x = F.relu(self.fc1(x)) x = self.fc2(x) return F.log_softmax(x) model = DataParallel(Net()) model.cuda() optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9) criterion = nn.NLLLoss().cuda() model.train() for batch_idx, (data, target) in enumerate(train_loader): input_var = Variable(data.cuda()) target_var = Variable(target.cuda()) print('Getting model output') output = model(input_var) print('Got model output') loss = criterion(output, target_var) optimizer.zero_grad() loss.backward() optimizer.step() print('Finished') ``` This doesn't throw an error, but hangs after it prints "Getting model output" and never returns. I traced this down to the `parallel_apply` spawning threads that then never finish. The line that hangs is [here](https://github.com/pytorch/pytorch/blob/e50a1f19b3dc735f0710929b97b0af384aafe09b/torch/nn/parallel/parallel_apply.py#L25) where the threads are spawned using both GPU 0 and GPU 1, but never finish. This is only a problem when `CUDA_VISIBLE_DEVICES=0,1` as both GPU0 and GPU1 work perfectly well individually. Before running this, `nvidia-smi` shows ``` +------------------------------------------------------+ | NVIDIA-SMI 352.68 Driver Version: 352.68 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla K80 Off | 0000:06:00.0 Off | 0 | | N/A 40C P0 57W / 149W | 55MiB / 11519MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla K80 Off | 0000:07:00.0 Off | 0 | | N/A 35C P0 76W / 149W | 55MiB / 11519MiB | 99% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+ ``` after running (while it hangs), `nvidia-smi` gives ``` +------------------------------------------------------+ | NVIDIA-SMI 352.68 Driver Version: 352.68 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla K80 Off | 0000:06:00.0 Off | 0 | | N/A 42C P0 69W / 149W | 251MiB / 11519MiB | 99% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla K80 Off | 0000:07:00.0 Off | 0 | | N/A 36C P0 90W / 149W | 249MiB / 11519MiB | 99% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 4785 C python 194MiB | | 1 4785 C python 192MiB | +-----------------------------------------------------------------------------+ ``` and `top` shows the main python process and the two python subprocesses. Wondering if this could be something similar to #554. Using [this](https://github.com/tensorflow/models/blob/master/tutorials/image/cifar10/cifar10_multi_gpu_train.py) TensorFlow example, I get linear speedup using multiple GPUs as I change `CUDA_VISIBLE_DEVICES` so multiple K80s should certainly be viable.

Are there any other solutions?

Thanks!

sabalaba · September 27, 2020, 4:46am

Hi! Because you’re a Lambda customer, can you please email support@lambdalabs.com with your issue? We should be able to fix it quickly.