[libfabric-users] ib latency question,can you help me?

Hefty, Sean sean.hefty at intel.com
Wed Apr 24 08:36:50 PDT 2019

> My HCA is ConnectX-3,and when I call ib_write_lat,what's the meaning of -C option?
> Because when I chose -C ,the t_avg of 1024 bytes is 8443 usec and 1GB is 2758537
> usec,it's too long and I don't understand the parameter -C(report-cycles report times
> in cpu cycle units) meaning.
> But when I cancel -C,the time becomes too short for 1024bytes etc,it's too fast.
> Can you solve this?

You should direct this question to linux-rdma at vger.kernel.org. 

It sounds like -C reports the number of CPU cycles, so 8443 would by cycles, not usec.  I'm not familiar with the output, so maybe that's a bug in the display.  The expected latency for that IB device is likely around 1 usec for a 1k write.

- Sean

More information about the Libfabric-users mailing list