cifar10 example is not scalable with multiple GPUs

[H4dr1en]

I slightly adapted the cifar10 example in this fork, basically removing python-fire and adding the torch.distributed.launch function, so that it can be executed as a standalone script with clearml-task.

I executed the following script with nproc_per_node in [1, 2, 3, 4] on a AWS g4dn.12xlarge instance (x4 T4 GPUs). I got the following results:

• batch size=16, nproc_per_node=1 => Runtime: 29:53
• batch size=16, nproc_per_node=1 => Runtime: 05:34
  Here I disabled DataParallel as mentionned in DataParallel is used by auto_model with single GPU pytorch/ignite#2447
• batch size=32, nproc_per_node=2 => Runtime: 17:11
• batch size=48, nproc_per_node=3 => Runtime: 11:33
• batch size=64, nproc_per_node=4 => Runtime: 08:47

I am increasing the batch size by 16 each time I add a GPU, so that each GPU has the same number of samples. I didn't change the default number of processes (8) for all of them, because I didn't oberserve that the GPUs were under-used (below 95%)

GPU utilization as reported by clearml

I was expecting to observe a quasi-linear time improvement, but it isn't the case. Am I missing something?

PS: Here are the requirements I used to execute the script

torch==1.7.1+cu110
torchvision==0.8.2
pytorch-ignite==0.4.8
clearml==.1.1.6
tensorboardX==2.4.1

[vfdev-5]

Thanks for reporting @H4dr1en ! I'd like to reproduce your results to see what happens once I have some time for that.

Today, we have some benchmarks for Pascal VOC on 1, 2 and 4 GPUs (GeForce RTX 2080 Ti)

• 1 GPU - 3:55h : https://app.clear.ml/projects/1f574b9622104ab0bcdef10c39ff5e2f/experiments/a971430efa76456895724ad7758bf44b/output/execution
• 2 GPUs - 1:53h : https://app.clear.ml/projects/1f574b9622104ab0bcdef10c39ff5e2f/experiments/5899d7044a9b4c3587f3b12ea498ba72/output/execution
• 4 GPUs - 1:02h : https://app.clear.ml/projects/1f574b9622104ab0bcdef10c39ff5e2f/experiments/f96e3ebd97104ab99ff94e582336afaf/output/execution

Btw, thanks for pointing out to clearml-task feature, cool feature !

[sdesrozis]

@H4dr1en Thanks for that report. It sounds weird, I had similar experiments when I worked in a research center with GPU cluster and it was fine regarding scalability.

Did you try disabling clearml ? Transferring the results to the server can create disruptions and interruptions during learning.

[H4dr1en]

Thank for your answers!

From what you reported, scaling the training speed linearly with the number of GPUs should be achievable, so there could be something wrong that could be fixed.

As a context, I do observe similar bad scalability for my own use case within g4dn.12xlarge instances, so I hope that if we can find the bottleneck with the cifar10 example, it would also unlock my other project.

[vfdev-5]

@H4dr1en can please you try original cifar10 example on your infrastructure using 1, 2, 4 GPUs and report back runtime here ?

CUDA_VISIBLE_DEVICES=0 python main.py run

# for older pytorch
python -u -m torch.distributed.launch --nproc_per_node=2 --use_env main.py run --backend="nccl"
# for >=1.9
torchrun --nproc_per_node=2 main.py run --backend="nccl"

# for older pytorch
python -u -m torch.distributed.launch --nproc_per_node=4 --use_env main.py run --backend="nccl"
# for >=1.9
torchrun --nproc_per_node=4 main.py run --backend="nccl"

my times on 1 and 2 GPUs GTX1080Ti to compare:

Config	Total time	Training epoch time, validation train/test times
1 GPU	00:04:22	00:00:08, 00:00:05, 00:00:02
2 GPUs (DDP)	00:02:57	00:00:06, 00:00:03, 00:00:01

[H4dr1en]

Thanks @vfdev-5 ! Here are my results:

Config	Total time	Training epoch time, validation train/test times
PyTorch 1.7.1+cu110, RTX 3070, 1 GPU	00:02:30	00:00:05, 00:00:01.8, 00:00:00.9
PyTorch 1.7.1+cu110, RTX 3070, 2 GPU	00:01:29	00:00:03, 00:00:01.3, 00:00:00.5
PyTorch 1.9.1+cu111, RTX 3070, 1 GPU	00:02:19	00:00:05, 00:00:01.8, 00:00:01
PyTorch 1.9.1+cu111, RTX 3070, 2 GPU	00:01:26	00:00:03, 00:00:01.3, 00:00:01
PyTorch 1.7.1+cu110, g4dn.12xlarge, 1 GPU (7Gb of GPU RAM used)	00:14:55	00:00:35, 00:00:02.4, 00:00:03
PyTorch 1.7.1+cu110, g4dn.12xlarge, 2 GPU	00:02:28	00:00:05, 00:00:01.7, 00:00:00.7
PyTorch 1.7.1+cu110, g4dn.12xlarge, 4 GPU	00:01:53	00:00:04, 00:00:01.7, 00:00:00.5

It looks like there is something wrong with 1 GPU on g4dn.12xlarge

Btw, is it fair to compare the speed this way? ie. in multi-gpu context, each GPU get a smaller batch_size

[vfdev-5]

Thanks for the results @H4dr1en !

Definitely, there is something unclear with 1 GPU case for g4dn.12xlarge.

Btw, is it fair to compare the speed this way? ie. in multi-gpu context, each GPU get a smaller batch_size

Well, I'd say we are interested how quickly the task was done. By the task we have to measure the number of processed images. If in multi-gpu context we load each GPU as in the a single GPU case, we have to reduce the number of iterations to run, otherwise they wont accomplish the same task, I think.

EDIT: in the logs for PyTorch 1.7.1+cu110, g4dn.12xlarge, 1 GPU (7Gb of GPU RAM used) case, do you see something like "Apply torch DataParallel on model" ?

[H4dr1en]

EDIT: in the logs for PyTorch 1.7.1+cu110, g4dn.12xlarge, 1 GPU (7Gb of GPU RAM used) case, do you see something like "Apply torch DataParallel on model" ?

No, here are the logs for this run:

ec2-user@ip-10-100-0-002:~/ignite/examples/contrib/cifar10# CUDA_VISIBLE_DEVICES=0 python main.py run
2022-02-04 12:22:46,263 ignite.distributed.launcher.Parallel INFO: - Run '<function training at 0x7f717e4a56a8>' in 1 processes
2022-02-04 12:22:49,529 CIFAR10-Training INFO: Train resnet18 on CIFAR10
2022-02-04 12:22:49,529 CIFAR10-Training INFO: - PyTorch version: 1.7.1+cu110
2022-02-04 12:22:49,529 CIFAR10-Training INFO: - Ignite version: 0.4.8
2022-02-04 12:22:49,536 CIFAR10-Training INFO: - GPU Device: Tesla T4
2022-02-04 12:22:49,536 CIFAR10-Training INFO: - CUDA version: 11.0
2022-02-04 12:22:49,536 CIFAR10-Training INFO: - CUDNN version: 8005
2022-02-04 12:22:49,536 CIFAR10-Training INFO:

2022-02-04 12:22:49,536 CIFAR10-Training INFO: Configuration:
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  with_amp: False
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  with_clearml: False
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  stop_iteration: None
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  nproc_per_node: None
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  log_every_iters: 15
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  resume_from: None
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  backend: None
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  checkpoint_every: 1000
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  validate_every: 3
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  num_warmup_epochs: 4
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  learning_rate: 0.4
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  num_epochs: 24
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  num_workers: 12
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  weight_decay: 0.0001
2022-02-04 12:22:49,536 CIFAR10-Training INFO:  momentum: 0.9
2022-02-04 12:22:49,537 CIFAR10-Training INFO:  batch_size: 512
2022-02-04 12:22:49,537 CIFAR10-Training INFO:  model: resnet18
2022-02-04 12:22:49,537 CIFAR10-Training INFO:  output_path: /tmp/output-cifar10/
2022-02-04 12:22:49,537 CIFAR10-Training INFO:  data_path: /tmp/cifar10
2022-02-04 12:22:49,537 CIFAR10-Training INFO:  seed: 543
2022-02-04 12:22:49,537 CIFAR10-Training INFO:

2022-02-04 12:22:49,537 CIFAR10-Training INFO: Output path: /tmp/output-cifar10/resnet18_backend-None-1_20220204-122249
Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to /tmp/cifar10/cifar-10-python.tar.gz
 99%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▉ | 169295872/170498071 [00:11<00:00, 15776620.49it/s]Extracting /tmp/cifar10/cifar-10-python.tar.gz to /tmp/cifar10
2022-02-04 12:23:04,002 ignite.distributed.auto.auto_dataloader INFO: Use data loader kwargs for dataset 'Dataset CIFAR10':
        {'batch_size': 512, 'num_workers': 12, 'shuffle': True, 'drop_last': True, 'pin_memory': True}
2022-02-04 12:23:04,003 ignite.distributed.auto.auto_dataloader INFO: Use data loader kwargs for dataset 'Dataset CIFAR10':
        {'batch_size': 1024, 'num_workers': 12, 'shuffle': False, 'pin_memory': True}
2022-02-04 12:23:10,650 CIFAR10-Training INFO: Engine run starting with max_epochs=24.

[H4dr1en]

But the behaviour above for 1 GPU on g4dn.12xlarge is probably a separate issue. Sorry I was not very explicit on the issue description, my main concern is the following:

If we define the factor of improvement as f(n_gpu) = training_time(n_gpu) / training_time(2 * n_gpu)

I would expect to have a factor of improvement approaching 2, and it seems to never be achieved. What could be the reason for that?

[vfdev-5]

I would expect to have a factor of improvement approaching 2, and it seems to never be achieved. What could be the reason for that?

I think in case of cifar10 a larger model can give something as a linear scaling (up to a certain limit).
EDIT: maybe dataset size and image size also play their role to that.
My times on 1 and 2 GPUs GTX1080Ti to compare for resnet152 model, 10 epochs

Config	Total time	Training epoch time, validation train/test times
1 GPU	00:05:21	00:00:25, 00:00:13, 00:00:03
2 GPUs (DDP)	00:03:51	00:00:19, 00:00:08, 00:00:02

See also results for Pascal VOC: #75 (comment) where factor ~ N GPUs for N = 1, 2 and 4