[Bug]Meet IB problem in single node experiment?

[TonyWu199]

Hi developers,
It is a really nice work and I try to reproduct in one node with 8 gpus. However, I meet the IB problems below in both UT and collective communication case.

The UT

command:
mpirun -np 2 ./test/mp_unit_tests

The all_reduce_test in c++

command:
mpirun --bind-to numa -np 8 ./test/mscclpp-test/allreduce_test_perf -b 3m -e 48m -G 100 -n 100 -w 20 -f 2 -k 5

In my view, the gpu communication intra a single node is unrelated to IB, right?
Could you help fix this problem or maybe some walkarounds?

[Binyang2014]

Please refer this comment: #254 (comment)