On Mon, Oct 06, 2008 at 12:22:40PM -0700, Aaron Turner wrote:
> On Mon, Oct 6, 2008 at 9:37 AM, Peter Van Epp <[EMAIL PROTECTED]> wrote:
> > <snip>
> >>
> >> Since he's using the memory cache and a pretty sizeable loop count,
> >> it's probably not disk I/O. I'm guessing most of it is overhead of
> >> doing 500K write()'s per second.
> >>
> >> When you're getting 995Mbps, I doubt your average packet size is 102
> >> bytes like Jeff's. If it is, then I'd love to hear more about your
> >> setup (kernel version, tuning, hardware, etc)
> >>
> >
> > You are correct, thats with 9K jumbo frames (these are HPC people,
> > they
> > don't do no stinkin small frames :-)). At 1500 it drops to about 650 megs.
> > The default untuned kernel 35 megabits was also 9k jumbo frames. Some time
> > I'll poke at why the drop at a 1500 MTU (perhaps something in tcp, perhaps
> > interrupt load perhaps something else). The bottom line is going fast
> > is hard with lots and lots of non obvious places that can bite you :-).
>
> Actually, I'd look into why you're only getting 995Mbps w/ jumbo frames:
>
> 650Mbps * 1024 * 1024 / 8bits/byte / 1514 bytes/packet = 56,272pps
>
> 995Mbps * 1024 * 1024 / 8bits/byte / 9014 bytes/packet = 14,468pps
>
> Honestly, neither of these numbers are that great IMHO... you should
> be able to do 100,000pps on most hardware nowadays.
>
> --
> Aaron Turner
> http://synfin.net/
> http://tcpreplay.synfin.net/ - Pcap editing and replay tools for Unix &
> Windows
As noted a gigabit light path, doing within 5 megabits per second
of line speed (plus I believe that netperf is reporting user data throughput
presumable leaving the extra as IP header overhead). There isn't any
more bandwidth to get, its running at capacity and the network is the
limitation (this was done a few years ago and the link has moved to 10 gigs
now). As noted we don't know why it slows down at a 1500 MTU and haven't done
any performance testing at 10 gigs so far. The 10 gig network isn't fully built
out yet so there aren't yet any other 10 gig endpoints only aggregations of
multiple gig links in to the file store currently. The host on this end is
a large AIX box running a 500 terabyte filestore so it can achieve the
same performance as iperf going to disk if the other end can sustain it which
they often can.
Peter Van Epp / Operations and Technical Support
Simon Fraser University, Burnaby, B.C. Canada
-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Tcpreplay-users mailing list
Tcpreplay-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/tcpreplay-users
Support Information: http://tcpreplay.synfin.net/trac/wiki/Support
