i would like to add the metric "net_bw" to NCCL tests in order to understand how well the Network interface was utilized.
How complicated would it be to figure out in nccl-tests which portion of the bus_bw used (NvLink/Shmem) and which used the network (TCP/IB) ?
i would like to add the metric "net_bw" to NCCL tests in order to understand how well the Network interface was utilized. How complicated would it be to figure out in nccl-tests which portion of the bus_bw used (NvLink/Shmem) and which used the network (TCP/IB) ?