'How to check the bandwidth usage of nccl backend?

I am using nccl backend to send and recv pytorch tensor on my server. And I am trying to use these command to check the bandwidth usage, is it corrent? Fisrt I use

'tcp://127.0.0.1:1224'

as my dist utils. And then I print

export NCCL_SOCKET_IFNAME=lo

lo is the network interface used for loop. Then I use

sudo iftop -i lo -n -P

But the bandwidth is too small. WHat should I do to check the bandwidth uses of dist.send and dist.recv?

linuxnetworkingpytorchdistributed-computingdistributed-system


Sources

This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.

Source: Stack Overflow

Solution Source

Related Questions