linux - 为什么 slurm 中的作业是 TensorFlow 脚本时会无限期卡住？

Question

我在使用 slurm ( http://slurm.schedmd.com/ ) 工作负载管理器时遇到此错误。当我运行一些 tensorflow python 脚本时，有时会导致错误(附件)。似乎找不到安装的 cuda 库，但我正在运行不需要 GPU 的脚本。因此，我很困惑为什么 cuda 会成为一个问题。如果我不需要 cuda 安装，为什么会出现问题？

我从 slurm-job_id 文件中获得的唯一有用信息如下:

I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:102] Couldn't open CUDA library libcudnn.so. LD_LIBRARY_PATH: /cm/shared/openmind/cuda/7.5/lib64:/cm/shared/openmind/cuda/7.5/lib
I tensorflow/stream_executor/cuda/cuda_dnn.cc:2092] Unable to load cuDNN DSO
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcuda.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcurand.so locally
E tensorflow/stream_executor/cuda/cuda_driver.cc:491] failed call to cuInit: CUDA_ERROR_NO_DEVICE
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:153] retrieving CUDA diagnostic information for host: node047
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:160] hostname: node047
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:185] libcuda reported version is: Not found: was unable to find libcuda.so DSO loaded into this program
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:347] driver version file contents: """NVRM version: NVIDIA UNIX x86_64 Kernel Module  352.63  Sat Nov  7 21:25:42 PST 2015
GCC version:  gcc version 4.8.5 20150623 (Red Hat 4.8.5-4) (GCC)
"""
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:189] kernel reported version is: 352.63.0
I tensorflow/core/common_runtime/gpu/gpu_init.cc:81] No GPU devices available on machine.

我一直认为 tensorflow 不需要 GPU。所以我假设最后一个错误说没有 GPU 不会导致错误(如果我错了请纠正我)。

我不明白为什么我需要 CUDA 库。我正在尝试使用 GPU 运行我的作业，如果我的作业是 CPU 作业，为什么我需要 cuda 库？

---

我尝试直接登录节点并启动 tensorflow，但没有出现明显的错误:

I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:102] Couldn't open CUDA library libcudnn.so. LD_LIBRARY_PATH: /cm/shared/openmind/cuda/7.5/lib64:/cm/shared/openmind/cuda/7.5/lib
I tensorflow/stream_executor/cuda/cuda_dnn.cc:2092] Unable to load cuDNN DSO
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcuda.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcurand.so locally

虽然我预料到错误:

I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:102] Couldn't open CUDA library libcudnn.so. LD_LIBRARY_PATH: /cm/shared/openmind/cuda/7.5/lib64:/cm/shared/openmind/cuda/7.5/lib
I tensorflow/stream_executor/cuda/cuda_dnn.cc:2092] Unable to load cuDNN DSO
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcuda.so locally
I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcurand.so locally
E tensorflow/stream_executor/cuda/cuda_driver.cc:491] failed call to cuInit: CUDA_ERROR_NO_DEVICE
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:153] retrieving CUDA diagnostic information for host: node047
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:160] hostname: node047
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:185] libcuda reported version is: Not found: was unable to find libcuda.so DSO loaded into this program
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:347] driver version file contents: """NVRM version: NVIDIA UNIX x86_64 Kernel Module  352.63  Sat Nov  7 21:25:42 PST 2015
GCC version:  gcc version 4.8.5 20150623 (Red Hat 4.8.5-4) (GCC)
"""
I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:189] kernel reported version is: 352.63.0
I tensorflow/core/common_runtime/gpu/gpu_init.cc:81] No GPU devices available on machine.

---

我还在tensorflow库中做了一个官方的git issue:

https://github.com/tensorflow/tensorflow/issues/3632

Answer 1

在通过批处理作业使用 slurm 提交的 tensorflow 中存在一些错误。

目前我通过在 slurm 上运行 srun 来绕过它。

在您的情况下，您还安装了 GPU 版本的 tensorflow 并在没有 GPU 的机器上运行它。这在您的案例中导致了另一个错误。