OpenMPI builds with CUDA and OpenCL dependency even with --with-cuda=no option

[PavelStishenko]

Background information

Tested on version 2.1.1 from source tarball.

---

Details of the problem

I tried to build OpenMPI from source with the following script:

shell$ ./configure --enable-mpi-thread-multiple CC=clang-3.9 CXX=clang++-3.9 --prefix=/home/user/openmpi2 --without-cuda  --with-cuda=no --disable-vt
shell$ make -j8 all
shell$ make -j8 install

Then I checked mpiexec dependecies with ldd command and found, that there are still dependencies on libOpenCL.so and libcudart.so. Of course on the server without these libraries this program do not work.

[jsquyres]

@sjeaugey Can you look into this?

[sjeaugey]

There is no reference to OpenCL in the CUDA-aware code. My only guess as to where those dependencies are coming from is hwloc. @bgoglin do you think it's possible that hwloc within Open MPI is compiling some CUDA and OpenCL support, creating this dependency ?

[bgoglin]

A quick look in opal/mca/hwloc/hwloc*/configure.m4 tells me that OMPI's configure doesn't prevent hwloc from enabling everything it finds except NVML.

[jsquyres]

Seems like a bug -- we should have the hwloc component honor the top-level Open MPI --without-cuda CLI option.

I'm guessing that this issue is present in most (all?) versions of Open MPI.

[bgoglin]

Yes, very likely present forever. Note that hwloc's OpenCL code is actually only useful for AMD GPU, not for NVIDIA. But it builds whenever OpenCL is detected.
Anyway, if OMPI doesn't care about OpenCL hwloc osdevs (and likely GL too), just apply something like this:

--- a/opal/mca/hwloc/hwloc2a/configure.m4
+++ b/opal/mca/hwloc/hwloc2a/configure.m4
@@ -122,6 +122,10 @@ AC_DEFUN([MCA_opal_hwloc_hwloc2a_CONFIG],[
     AS_IF([test "$opal_check_cuda_happy" != "yes"],
           [enable_nvml=no])
 
+    # GL and OpenCL osdevs aren't used in OMPI
+    enable_gl=no
+    enable_opencl=no
+
     # hwloc checks for compiler visibility, and its needs to do
     # this without "picky" flags.
     opal_hwloc_hwloc2a_save_cflags=$CFLAGS

[jsquyres]

@bgoglin Have a look at #4245 -- is your solution or my solution better? I'm happy to update #4245 with your solution if it's better.

[bgoglin]

I'd say apply both. I don't think your patch will disable OpenCL because hwloc's configure options for CUDA and OpenCL are independent.

[jsquyres]

@bgoglin Oh, good point -- I missed that distinction (CL vs. CUDA). Ok, I'll add your patch to #4245; thanks.

[jsquyres]

#4245 updated.

[sjeaugey]

@jsquyres @bgoglin 👍 this is exactly what I was looking for, just didn't find where to cleanly set those variables.

In general, we should only build what we're using. So we shouldn't need CUDA support in hwloc (inside Open MPI). That could apply to anything hwloc has support for.