10

I am new with TensorFlow.

I just installed TensorFlow and to test the installation, I tried the following code and as soon as I initiate the TF Session, I am getting the Segmentation fault (core dumped) error.

bafhf@remote-server:~$ python
Python 3.6.5 |Anaconda, Inc.| (default, Apr 29 2018, 16:14:56) 
[GCC 7.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
/home/bafhf/anaconda3/envs/ismll/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
>>> tf.Session()
2018-05-15 12:04:15.461361: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1349] Found device 0 with properties: 
name: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235
pciBusID: 0000:04:00.0
totalMemory: 11.17GiB freeMemory: 11.10GiB
Segmentation fault (core dumped)

My nvidia-smi is:

Tue May 15 12:12:26 2018       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.30                 Driver Version: 390.30                    |
|-------------------------------+----------------------+----------------------+
| 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           On   | 00000000:04:00.0 Off |                    0 |
| N/A   38C    P8    26W / 149W |      0MiB / 11441MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla K80           On   | 00000000:05:00.0 Off |                    2 |
| N/A   31C    P8    29W / 149W |      0MiB / 11441MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

And nvcc --version is:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176

Also gcc --version is:

gcc (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 20160609
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Following is my PATH:

/home/bafhf/bin:/home/bafhf/.local/bin:/usr/local/cuda/bin:/usr/local/cuda/lib:/usr/local/cuda/extras/CUPTI/lib:/home/bafhf/anaconda3/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin

and the LD_LIBRARY_PATH:

/usr/local/cuda/bin:/usr/local/cuda/lib:/usr/local/cuda/extras/CUPTI/lib


I am running this on a server and I don't have root privileges. Still I managed to install everything as per the instructions on the official website.

Edit: New observations:

Seems like the GPU is allocating memory for the process for a second and then the core segmentation dumped error is thrown:

Terminal output

Edit2: Changed tensorflow version

I downgraded my tensorflow version from v1.8 to v1.5. The issue still remains.


Is there any way address or debug this issue?

Benison Sam
  • 2,755
  • 7
  • 30
  • 40
  • Okay I am sorry this is more of a brain dump than anything else, I do not have the time right now to get into this more. I remember I had a similar issue once and since than my inner alarm bells ring when I see CUDA 9 with GCC > 6. I am sorry if that is a completely wrong clue, but maybe it has something to do with that. Also I personally do not know Anaconda very well, since I don't use it. – Alexander Harnisch May 15 '18 at 11:01
  • I updated my GCC version in the question and its 5.4. Will that be an issue? – Benison Sam May 15 '18 at 11:11
  • That should be fine as far as I know. – Alexander Harnisch May 15 '18 at 11:32
  • You should install CUDA Samples and check that CUDA's working. – MWB May 20 '18 at 23:11
  • 1
    Please refer https://github.com/tensorflow/tensorflow/issues/5934 – Sreeragh A R May 21 '18 at 12:21
  • @SreeraghAR, Do you want me to set the environment variable CUDA_VISIBLE_DEVICES? Because that's the gist of what I got from the solution given there.. – Benison Sam May 21 '18 at 13:05
  • Yes. This problem seems to be similar to yours. Please see if it helps – Sreeragh A R May 21 '18 at 13:30

7 Answers7

1

This could possibly occur since you are using multiple GPUs here. Try setting cuda visible devices to just one of the GPUs. See this linkfor instructions on how to do that. In my case, this solved the problem.

optimist
  • 1,018
  • 13
  • 26
  • 1
    I already did this and it works for the first GPU, but not with the second one and also not together. What might be causing it not to work with the second GPU? – Benison Sam May 23 '18 at 19:43
1

If you can see the nvidia-smi output, the second GPU has an ECC code of 2. This error manifests itself irrespective of a CUDA version or TF version error, and usually as a segfault, and sometimes, with the CUDA_ERROR_ECC_UNCORRECTABLE flag in the stack trace.

I got to this conclusion from this post:

"Uncorrectable ECC error" usually refers to a hardware failure. ECC is Error Correcting Code, a means to detect and correct errors in bits stored in RAM. A stray cosmic ray can disrupt one bit stored in RAM every once in a great while, but "uncorrectable ECC error" indicates that several bits are coming out of RAM storage "wrong" - too many for the ECC to recover the original bit values.

This could mean that you have a bad or marginal RAM cell in your GPU device memory.

Marginal circuits of any kind may not fail 100%, but are more likely to fail under the stress of heavy use - and associated rise in temperature.

A reboot usually is supposed to take away the ECC error. If not, seems like the only option is to change the hardware.

So what all I did and finally how I fixed the issue?

  1. I tested my code a on a separate machcine with NVIDIA 1050 Ti machine and my code executed perfectly fine.
  2. I made the code run only on the first card for which the ECC value was normal, just to narrow down the issue. This I did following, this post, setting the CUDA_VISIBLE_DEVICES environment variable.

  3. I then requested for restart of the Tesla-K80 server to check whether a restart can fix this issue, they took a while but the server was then restarted

    Now the issue is no more and I can run both the cards for my tensorflow implemntations.

Benison Sam
  • 2,755
  • 7
  • 30
  • 40
1

In case anyone still interested in, I happened to had the same issue, with "Volatile Uncorr. ECC" output. My problem was incompatible versions as shown below:

Loaded runtime CuDNN library: 7.1.1 but source was compiled with: 7.2.1. CuDNN library major and minor version needs to match or have higher minor version in case of CuDNN 7.0 or later version. If using a binary install, upgrade your CuDNN library. If building from sources, make sure the library loaded at runtime is compatible with the version specified during compile configuration. Segmentation fault

After I upgrade CuDNN library to 7.3.1 (which is greater than 7.2.1), segmentation fault error disappeared. To upgrade I did the following (as also documented in here).

  1. Download CuDNN library from NVIDIA website
  2. sudo tar -xzvf [TAR_FILE]
  3. sudo cp cuda/include/cudnn.h /usr/local/cuda/include
  4. sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
  5. sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
fatih
  • 1,171
  • 2
  • 14
  • 26
1

I was also facing the same issue. I have a workaround for the same you can try that.

I followed the following steps: 1. Reinstall the python 3.5 or above 2. Reinstall the Cuda and Add the Cudnn libraries to it. 3. Reinstall Tensorflow 1.8.0 GPU version.

vineagle
  • 39
  • 6
0

Check that you are using the exact version of CUDA and CuDNN required by tensorflow, and also that you are using the version of driver of the graphics card that comes with this CUDA version.

I once had a similar issue having a driver that was too recent. Downgrading it to the version coming with the CUDA version required by tensorflow solved the issue for me.

P-Gn
  • 23,115
  • 9
  • 87
  • 104
0

I encounter this problem recently.

The reason is multiple GPUs in docker container. The solution is pretty simple, you either:

set CUDA_VISIBLE_DEVICES in host refers to https://stackoverflow.com/a/50464695/2091555

or

use --ipc=host to launch the docker if you need multiple GPUs e.g.

docker run --runtime nvidia --ipc host \
  --rm -it
  nvidia/cuda:10.0-cudnn7-runtime-ubuntu16.04:latest

This problem is actually pretty nasty, and segfault happens during cuInit() calls in docker container and everything works fine in the host. I will leave log here to let the search engine find this answer easier for other people.

(base) root@e121c445c1eb:~# conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
Collecting package metadata (current_repodata.json): / Segmentation fault (core dumped)

(base) root@e121c445c1eb:~# gdb python /data/corefiles/core.conda.572.1569384636
GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.5) 7.11.1
Copyright (C) 2016 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from python...done.

warning: core file may not match specified executable file.
[New LWP 572]
[New LWP 576]

warning: Unexpected size of section `.reg-xstate/572' in core file.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Core was generated by `/opt/conda/bin/python /opt/conda/bin/conda upgrade conda'.
Program terminated with signal SIGSEGV, Segmentation fault.

warning: Unexpected size of section `.reg-xstate/572' in core file.
#0  0x00007f829f0a55fb in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so
[Current thread is 1 (Thread 0x7f82bbfd7700 (LWP 572))]
(gdb) bt
#0  0x00007f829f0a55fb in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so
#1  0x00007f829f06e3a5 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so
#2  0x00007f829f07002c in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so
#3  0x00007f829f0e04f7 in cuInit () from /usr/lib/x86_64-linux-gnu/libcuda.so
#4  0x00007f82b99a1ec0 in ffi_call_unix64 () from /opt/conda/lib/python3.7/lib-dynload/../../libffi.so.6
#5  0x00007f82b99a187d in ffi_call () from /opt/conda/lib/python3.7/lib-dynload/../../libffi.so.6
#6  0x00007f82b9bb7f7e in _call_function_pointer (argcount=1, resmem=0x7ffded858980, restype=<optimized out>, atypes=0x7ffded858940, avalues=0x7ffded858960, pProc=0x7f829f0e0380 <cuInit>, 
    flags=4353) at /usr/local/src/conda/python-3.7.3/Modules/_ctypes/callproc.c:827
#7  _ctypes_callproc () at /usr/local/src/conda/python-3.7.3/Modules/_ctypes/callproc.c:1184
#8  0x00007f82b9bb89b4 in PyCFuncPtr_call () at /usr/local/src/conda/python-3.7.3/Modules/_ctypes/_ctypes.c:3969
#9  0x000055c05db9bd2b in _PyObject_FastCallKeywords () at /tmp/build/80754af9/python_1553721932202/work/Objects/call.c:199
#10 0x000055c05dbf7026 in call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>) at /tmp/build/80754af9/python_1553721932202/work/Python/ceval.c:4619
#11 _PyEval_EvalFrameDefault () at /tmp/build/80754af9/python_1553721932202/work/Python/ceval.c:3124
#12 0x000055c05db9a79b in function_code_fastcall (globals=<optimized out>, nargs=0, args=<optimized out>, co=<optimized out>)
    at /tmp/build/80754af9/python_1553721932202/work/Objects/call.c:283
#13 _PyFunction_FastCallKeywords () at /tmp/build/80754af9/python_1553721932202/work/Objects/call.c:408
#14 0x000055c05dbf2846 in call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>) at /tmp/build/80754af9/python_1553721932202/work/Python/ceval.c:4616
#15 _PyEval_EvalFrameDefault () at /tmp/build/80754af9/python_1553721932202/work/Python/ceval.c:3124
... (stack omitted)
#46 0x000055c05db9aa27 in _PyFunction_FastCallKeywords () at /tmp/build/80754af9/python_1553721932202/work/Objects/call.c:433
---Type <return> to continue, or q <return> to quit---q
Quit

Another try is using pip to install

(base) root@e121c445c1eb:~# pip install torch torchvision
(base) root@e121c445c1eb:~# python
Python 3.7.3 (default, Mar 27 2019, 22:11:17) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
Segmentation fault (core dumped)

(base) root@e121c445c1eb:~# gdb python /data/corefiles/core.python.28.1569385311 
GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.5) 7.11.1
Copyright (C) 2016 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from python...done.

warning: core file may not match specified executable file.
[New LWP 28]

warning: Unexpected size of section `.reg-xstate/28' in core file.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
bt
Core was generated by `python'.
Program terminated with signal SIGSEGV, Segmentation fault.

warning: Unexpected size of section `.reg-xstate/28' in core file.
#0  0x00007ffaa1d995fb in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
(gdb) bt
#0  0x00007ffaa1d995fb in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#1  0x00007ffaa1d623a5 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#2  0x00007ffaa1d6402c in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#3  0x00007ffaa1dd44f7 in cuInit () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#4  0x00007ffaee75f724 in cudart::globalState::loadDriverInternal() () from /opt/conda/lib/python3.7/site-packages/torch/lib/libtorch_python.so
#5  0x00007ffaee760643 in cudart::__loadDriverInternalUtil() () from /opt/conda/lib/python3.7/site-packages/torch/lib/libtorch_python.so
#6  0x00007ffafe2cda99 in __pthread_once_slow (once_control=0x7ffaeebe2cb0 <cudart::globalState::loadDriver()::loadDriverControl>, 
... (stack omitted)
Cloud
  • 1,374
  • 1
  • 13
  • 26
-1

I am using tensorflow in a cloud enviornment from paperspace.

Update of cuDNN 7.3.1 did not work for me.

One way is to build Tensorflow with proper GPU and CPU support.

This is not proper solution but this solved my issue temporarily (downgrade tensoflow to 1.5.0):

pip uninstall tensorflow-gpu
pip install tensorflow==1.5.0
pip install numpy==1.14.0
pip install six==1.10.0
pip install joblib==0.12

Hope this helps !

morgan121
  • 2,213
  • 1
  • 15
  • 33
Dwijay Bane
  • 110
  • 6
  • I believe this solution doesn't address the issue described or maybe is a specific solution only for your scenario – Benison Sam Jan 07 '19 at 21:59