Re: Peering Traffic Volume

["Patrick W. Gilmore" <patrick () ianai net>]

On Mar 25, 2011, at 1:44 PM, Bill Woodcock wrote:
> On Mar 24, 2011, at 4:27 PM, Ravi Ramaswamy wrote:
> >

> > I am using 2.5 Tbps as the peak volume of peering traffic over all peering
> > points for a Tier 1 ISP, for some modeling purposes.  Is that a reasonable
> > estimate?
>
> That's actually a very difficult research question for the academic community, and one that they've been struggling 
> with since they lost their overview of the NSFNET backbone in ~1992.
>
> Ironically, it's quite easy for any one ISP to answer internally, but these numbers are closely held as trade-secrets.
>
> One thing you can do is look at the total volume of publicly-reported traffic across IXP switch fabrics:
>
>    https://prefix.pch.net/applications/ixpdir/summary/growth-region/?sort1=bandwidth&sort2=_current&order=desc 
>
>    https://prefix.pch.net/applications/ixpdir/?show_active_only=0&sort=traffic&order=desc
>
> …where you see about 8.3Tbps of overall reported traffic.  Then you could do various analyses comparing IXPs where 
> crossconnects are prevalent (Equinix Ashburn, say) to ones where they are not, and looking at which ISPs peer at 
> each.  You could also try to find out from ISPs which IXPs they use crossconnects at, and which they don't.  That may 
> be easier information for you to get than how much traffic they're doing individually.

IXP vs. private interconnect (be it peering or customer/transit) ratios varies dramatically with geography, scale, and 
even the proclivities of the various network architects.

The question is whether "some data" is better than "no data".  Honestly, I'm not sure.  I see lots of things with 'some 
data' that are actually worse than guesses.  But where I cannot eventually find the actual answer, I am left wanting to 
prefer the "some data" ones, probably because "data" sounds good.


> It might also be interesting to look at some of the IXPs that publish per-participant traffic figures, to see if you 
> can develop characteristic statistical distributions for amount-each-participant-contributes-to-the-IXP, though you 
> should be cautioned that the curve might be much heavier-tailed for a large exchange than a small one.

Even that is dangerous.  For instance, some participants move traffic away from such IXPs.  (That is not a guess, I 
know first hand that this happens - especially as I am one of those people.)

-- 
TTFN,
patrick