Hi, On Mon, Sep 13, 2010 at 10:18:50AM +0100, Nick Hilliard wrote: > On 13/09/2010 07:05, Gert Doering wrote: > >ports, while the average egress load never exceeded 50% (!) > > The average that you're talking about here is measured over 5 minutes, > which is an eternity in terms of packet throughput. If you drop your > measurement interval from 5 minutes to something much lower (5 - 30 > seconds, you'll get a much more accurate picture of what your switch is > actually doing.
Nick, grant me a bit of understanding about averaging and bursts :-) In our case, the workload was "audio streaming servers" on the ingress and "port to the router" on the egress. We knew the number of users per server, we know the traffic profile per stream, so we have a fairly good understanding of the traffic averaged per second. Now the nasty bit is how these servers distribute the packets they send out in each individual second - Windows Media does this fairly "nicely" (spreading out the packets), while most other streaming software creates somewhat massive wirespeed bursts, and then waits some milliseconds, and then generates a new wirespeed burst. With 2970s, no drops on egress. With 2960G, hell on earth, drops start at about 25% average load on the egress port, and increase to "the users complain that the audio is inacceptable" at about 50% average load. We are now using a 4x GigE channel with an average load of below 1 Gbit/s as "egress port", and *that* can now keep up with the "load" with very little packet drops. But we're in the lucky situation that this "just" causes some cabling effort and +3 ports on the upstream 6500, which was much less expensive than "unhappy customer". gert -- USENET is *not* the non-clickable part of WWW! //www.muc.de/~gert/ Gert Doering - Munich, Germany g...@greenie.muc.de fax: +49-89-35655025 g...@net.informatik.tu-muenchen.de
pgpJgRJy9b4V6.pgp
Description: PGP signature
_______________________________________________ cisco-nsp mailing list cisco-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/cisco-nsp archive at http://puck.nether.net/pipermail/cisco-nsp/