I've seen a similar issue when libcuda.so is not in the LD_LIBRARY_PATH (on
one of my systems, only libcuda.so.1 was there, but the usual libcuda.so
symlink was absent). Can you please check that? If it turns out that's
your issue, you can either simply ln -s libcuda.so.1 libcuda.so in the
relevant directory or you can modify NCCL to dlopen libcuda.so.1 instead of
libcuda.so.

Thanks,
Cliff

On Thu, Jan 14, 2016 at 10:28 PM, Jerry Lin [email protected]
wrote:

I build nccl with cuda-7.5:

make CUDA_HOME=/usr/local/cuda-7.5 test

And run test with the following command:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:./build/lib
./build/test/all_reduce_test

causes a segmentation fault:

Using devices

Segmentation fault

But all tests run smoothly if I build nccl with cuda-7.0.
Is the current version of nccl not compatible with cuda-7.5?

—
Reply to this email directly or view it on GitHub
#8.

from nccl.

Oh, and yes, NCCL is normally compatible with CUDA 7.5. It actually is a
bit more complete on CUDA 7.5 than on 7.0, since 7.0 lacked some of the
necessary support for the fp16 'half' datatype.

On Thu, Jan 14, 2016 at 10:41 PM, Cliff Woolley [email protected]
wrote:

I've seen a similar issue when libcuda.so is not in the LD_LIBRARY_PATH
(on one of my systems, only libcuda.so.1 was there, but the usual
libcuda.so symlink was absent). Can you please check that? If it turns
out that's your issue, you can either simply ln -s libcuda.so.1 libcuda.so in the relevant directory or you can modify NCCL to dlopen
libcuda.so.1 instead of libcuda.so.

Thanks,
Cliff

On Thu, Jan 14, 2016 at 10:28 PM, Jerry Lin [email protected]
wrote:

I build nccl with cuda-7.5:

make CUDA_HOME=/usr/local/cuda-7.5 test

And run test with the following command:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:./build/lib
./build/test/all_reduce_test

causes a segmentation fault:

Using devices

Segmentation fault

But all tests run smoothly if I build nccl with cuda-7.0.
Is the current version of nccl not compatible with cuda-7.5?

—
Reply to this email directly or view it on GitHub
#8.

from nccl.

Side note: assuming this is the same issue with libcuda.so that I'm referring to, we should fix the tests to fail more gracefully when the communicator cannot be created. The segfault happens when we pass a NULL communicator to some subsequent routine.

from nccl.

@cliffwoolley Thanks for the explanation.
I create a symlink to libcuda.so.1 and now it works!
So it's the same issue with libcuda.so.

from nccl.

Great! Glad to hear it. We'll leave this issue open to deal with the libcuda.so[.1] loading (perhaps we could try both variants before giving up) as well as to detect communicator creation failure in the test apps without segfaulting. I believe @nluehr already has fixes pending for one or both of these issues.

from nccl.

These issues are resolved in change sets caa40b8 and 2758353.

from nccl.