環境
- 4 GeForce GTX 1080 GPUS
- docker image
nnabla/nnabla-ext-cuda-multi-gpu:py36-cuda102-mpi3.1.6-v1.14.0
代碼
- 從倉庫nnabla-ext-cuda-multi-gpu拉取鏡像
docker pull nnabla/nnabla-ext-cuda-multi-gpu:py36-cuda102-mpi3.1.6-v1.14.0 - 運行
docker run -it --rm --gpus all nnabla/nnabla-ext-cuda-multi-gpu:py36-cuda102-mpi3.1.6-v1.14.0 - 添加
test.py
import nnabla.communicators as C
from nnabla.ext_utils import get_extension_context
extension_module = "cudnn"
ctx = get_extension_context(extension_module)
comm = C.MultiProcessCommunicator(ctx)
comm.init()
print(f'sizes={comm.size}, divice_id={comm.rank}')
- 運行
mpiexec -np 4 python test.py將會拋出異常。（異常只發生在使用GPU數大於2時）
bug
拋出異常如下：
Traceback (most recent call last):
File "test.py", line 6, in <module>
comm.init()
File "communicator.pyx", line 121, in nnabla.communicator.Communicator.init
RuntimeError: target_specific error in init
/home/gitlab-runner/builds/g9zRZKFe/2/nnabla/builders/all/nnabla-ext-cuda/src/nbla/cuda/communicator/multi_process_data_parallel_communicator.cu:358
ncclCommInitRank failed.
使用NCCL_DEBUG=INFO查看詳細信息mpiexec -np 4 -x NCCL_DEBUG=INFO python test.py
...
0db89117f3b2:87:87 [2] include/shm.h:28 NCCL WARN Call to posix_fallocate failed : No space left on device
0db89117f3b2:87:87 [2] NCCL INFO include/shm.h:41 -> 2
0db89117f3b2:87:87 [2] include/shm.h:48 NCCL WARN Error while creating shared memory segment nccl-shm-recv-6d2dacd576938b74-0-3-2 (size 9637888)
...
可以看到沒有多余的共享內存，但是使用nvidia-smi查看GPU情況，發現內存並沒有過多使用。
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.36.06 Driver Version: 450.36.06 CUDA Version: 11.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 GeForce GTX 1080 On | 00000000:01:00.0 Off | N/A |
| 27% 30C P8 5W / 180W | 815MiB / 8119MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 GeForce GTX 1080 On | 00000000:02:00.0 Off | N/A |
| 27% 33C P8 6W / 180W | 4MiB / 8119MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 2 GeForce GTX 1080 On | 00000000:03:00.0 Off | N/A |
| 28% 35C P8 5W / 180W | 4MiB / 8119MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 3 GeForce GTX 1080 On | 00000000:04:00.0 On | N/A |
| 28% 34C P8 6W / 180W | 4MiB / 8118MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
+-----------------------------------------------------------------------------+
原因
異常原因是NCCL不能在/dev/shm創建共享內存文件。因為docker默認的/dev/shm文件的大小為64MB太小了，所以當使用GPU數大於2時，會顯示內存不夠。
解決
有3中方式：
- 在
/etc/nccl.conf或~/.nncd.conf文件中，添加配置NCCL_SHM_DISABLE=1；如果nccl版本小於2.7，還應該配置NCCL_P2P_LEVEL=SYS，具體參考NCCL配置。 - 可以映射宿主機上的
/dev/shm，即docker run -v /dev/shm:/dev/shm ...，但是這樣會在宿主機上留下臟文件。 - 運行時，修改容器共享內存的大小，即
docker run --shm-size=256m ...。