[openib-general] ia64 perf and FMR
Grant Grundler
iod00d at hp.com
Mon Apr 4 12:20:03 PDT 2005
On Mon, Apr 04, 2005 at 08:20:02AM -0700, Libor Michalek wrote:
> Using libsdp with netperf shows some of the limitations of libsdp.
> netperf connects to the server on a well known socket, but then the
> server creates a second socket which it autobinds, checks to see
> which port was assigned, and passes the port number to the client
> which connects to the port. This second connection is then used for
> the data transfer.
Ah ok. That explains why just keying off the port # didn't work.
I wasn't aware of that.
> Since the connection is not on a well known port the way to match
> it is with the 'program' keyword:
>
> match program netperf
Wouldn't we also need a line like this?
match program netserver
For a sanity check, I sometimes run netperf first in one direction
and then the other. It helps confirm the two boxes are symetrical
(same HW config, same CPU, same firmware, same drivers, etc).
Having one libsdp.conf would keep things easy.
And in fact, my current configuration is NOT symetrical.
I have the same HW config but system firmware is not the same.
This results in ~8-10% loss in performance in one direction vs the other.
> > I'm now getting ~5300-5500 Mb/s (~660 MB/s) using SDP with netperf.
> > (256KB socket size....probably too small).
>
> The socket size socket option is still in the TODO file.
OH! I was working out how long it takes IB card to transmit or fill a
256KB buffer and it's really not very long. Kind of limits how many
transactions can be coalesced into one interrupt and how long the
interrupt handler can be deferred. But that's probably not a burning
issue (yet).
thanks,
grant
More information about the general
mailing list