[j-nsp] EX Routing Throughput

[Richard A Steenbergen]

On Sun, Nov 08, 2009 at 10:09:12PM +0100, Benny Amorsen wrote:
> Richard A Steenbergen <ras at e-gerbil.net> writes:
> 
> > On Sun, Nov 08, 2009 at 04:47:35PM +0200, Eugeniu Patrascu wrote:
> >> Maybe a bit late, but this is the most I saw on an EX4200 on 10GE
> >> ports under test.
> > ...
> >>   Input rate    ?: 7618850816 bps (14880568 pps)
> >>   Output rate   ?: 7618851328 bps (14880569 pps)
> >
> > Good, that is line rate (or at least really, really close). See this 
> > post about Ethernet overhead:
> >
> > http://www.mail-archive.com/juniper-nsp@puck.nether.net/msg06957.html
> 
> Both number divide to exactly 512 byte payload. Let us then assume 38
> byte overhead per packet, and the result is:
> 
> 8184312400 bps input and 8184312950 output. Not particularly close to
> line rate.

I think you're getting your numbers confused. 7618850816 bits/sec / 8
bits/byte = 952356352 bytes/sec / 14880568 packets/sec = 64.000
bytes/packet, which is precisely what we'd expect here. The other 20
bytes of overhead are pure layer 1 and won't be shown in the counters,
SFD/preamble for 8 bytes, IFG for 12 bytes, giving you a total of 84
bytes/packet consumed wire time. 84 * 14880568 = 1249967712 bytes/sec *
8 = 9999741696 bps, which like I said is really really close (and
probably just an error in measurement or rate calculation). Even if you
can't figure out that its line rate based on the bps, the dead giveaway 
is the pps. The theoretical max pps rate on 10GE with no additional 
overhead (no mpls, vlan tags, etc) is 14,880,952 pps, which is almost 
precisely what is being reported above.

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)