fix(build): check for IB verbs extension(s) support

[miEsMar]

Description

Latest NCCL implementations added support for
Infiniband extension features, such as active_speed_ex in struct ibv_port_attr. However, such additional support was not added in a backward compatible way.
This fix adds proper configuration-time checks to make sure the Infiniband installation supports newer features and extensions, instead of blindly assuming they are supported,
causing compile errors in the "net_ib" transport component.

Credits: AccelCom @ Barcelona Supercomputing Center

Related Issues

N/A.

Changes & Impact

N/A.

Performance Impact

N/A.

[sjeaugey]

I'm not sure I understand what problem we're solving here.

Under which conditions do you see compilation errors? Can you provide the build command line and the errors in the output ?

[miEsMar]

If NCCL is built with RDMA_CORE=1, and the system supports an IB version older than the one(s) which included support for extensions, then compilation fails since NCCL blindly assumes that they are always supported.

Build command:

    make src.${_make_tgt} \
        VERBOSE=1 \
        TRACE=1 \
        RDMA_CORE=1 \
        NET_PROFILER=1 \
        CC=gcc \
        CXX=g++ \
        PREFIX=${INSTALLDIR} \
        CUDA_HOME=${NVCUDA_ROOT} \
        NVCC_GENCODE=${_gpu_archs} \
        -j ${_nthreads}

Error:

transport/net_ib/init.cc:357:28: error: 'struct ibv_port_attr' has no member named 'active_speed_ex'; did you mean 'active_speed'?
  357 |               if (portAttr.active_speed_ex) {
      |                            ^~~~~~~~~~~~~~~
      |                            active_speed
transport/net_ib/init.cc:359:70: error: 'struct ibv_port_attr' has no member named 'active_speed_ex'; did you mean 'active_speed'?
  359 |                 ncclIbDevs[ncclNIbDevs].speed = ncclIbSpeed(portAttr.active_speed_ex) * ncclIbWidth(portAttr.active_width);
      |                                                                      ^~~~~~~~~~~~~~~
      |                                                                      active_speed

Therefore, the PR adds a check at build time that preemptively checks for actual IB extensions supports, and only includes usage if they are, otherwise falls back to base features.