nv_bandwidth

prerequests:

apt-get install libnccl2 libnccl-dev

compile:

nvcc -O3 -std=c++17 -c bw_test_multi.cu -o bw_test.o   -I/usr/local/openmpi/include
/usr/local/openmpi/bin/mpicxx -O3 bw_test.o -o exp_baseline   -L/usr/local/cuda-12.2/targets/x86_64-linux/lib -lcudart   -lnuma -lnccl

run:

# args: mode min_mib max_mib iters
CUDA_VISIBLE_DEVICES=1  mpirun -n 1 --allow-run-as-root ./exp_baseline 0 0.25 512 100

mode:

mode	test case
0	Single GPU load pinned memory
1	Multi GPU load same share memory
2	Multi GPU load different pinned memory
3	Pinned → GPU0 → NCCL broadcast to other GPUs