UCX warnings for GPU complex_double tests with OpenMPI
Reproducer:
module purge
module load cuda/11.4 gcc/11 openmpi/4 mkl/2022.1 nccl/2.11.4
export OMPI_MCA_coll=^hcoll
../configure --prefix=$HOME/soft/elpa_mpi_00 --enable-option-checking=fatal CC=mpicc FC=mpif90 CXX=mpicxx CFLAGS="-O3 -g -march=skylake-avx512 -I$MKL_HOME/include/intel64/lp64 -I$CUDA_HOME/include" CXXFLAGS="-std=c++17 -O3 -march=skylake-avx512 -I$MKL_HOME/include/intel64/lp64 -I$CUDA_HOME/include" FCFLAGS="-O3 -g -march=skylake-avx512 -I$MKL_HOME/include/intel64/lp64 -I$CUDA_HOME/include" LDFLAGS="-L$MKL_HOME/lib/intel64 -lmkl_scalapack_lp64 -lmkl_gf_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_openmpi_lp64 -lpthread -Wl,-rpath,$MKL_HOME/lib/intel64" --with-mpi=yes --enable-assumed-size --enable-band-to-full-blocking --enable-nvidia-gpu --with-NVIDIA-GPU-compute-capability=sm_70 -with-cuda-path=$CUDA_HOME --enable-avx512 --enable-cpp-tests=no --enable-single-precision --enable-nvtx
The warnings like
[1702644215.666499] [ravg1002:132812:0] mpool.c:55 UCX WARN object 0xcf82c0 {{cpml|cb|snd_tag|rk_use} send length 41943040 ucp_proto_progress_tag_rndv_rts() comp:mca_pml_ucx_send_nbx_completion()host me was not returned to mpool ucp_requests
appear for complex_double tests, e.g. validate_complex_double_eigenvectors_1stage_gpu_random
but not for real_double