[openib-general] questions and a comment on the perftest package

Michael S. Tsirkin mst at mellanox.co.il
Tue Nov 7 02:38:33 PST 2006


Quoting r. Or Gerlitz <ogerlitz at voltaire.com>:
> Subject: Re: questions and a comment on the perftest package
> 
> Michael S. Tsirkin wrote:
> > Quoting Or Gerlitz <ogerlitz at voltaire.com>:
> >> Whouldn't it make more sense to get one time stamp before the i'th posting
> >> and one tstamp after the i'th completion is reaped from the cq?
> > 
> > That's what we do, anyway - look how this works:
> > 
> > stamp[0]
> > post
> > poll
> > stamp[1]
> > post
> > poll
> > stamp[2]
> > ....
> > 
> > so stamp[i] is taken before the i'th posting
> > and stamp[i+1] is after the i'th completion.
> 
> oops, i have just noted that read_lat.c practically ignores the tx_depth 
> param... so stamp[i+1]-stamp[i] is indeed the wall time of the i'th 
> operation. Anyway, i guess you would be open to get a patch that does 
> exercise tx_depth in a similar fashion to read_bw.c ?

Something like this has been on my todo list for a while.
However, isn't it the case that just giving tx depth = 1 to rdma_bw we get
all the necessary deltas?

So the right thing to do, IMO, is to take rdma_bw and teach it to report latency
as well.  We thus will have a single test that measures both BW and latency for
reads, and have number of in-flight messages as parameter.  With tx_depth = 1
we'll get ping-pong.

-- 
MST




More information about the general mailing list