Interesting numbers. After a few weeks of bad performance on an Internet feed I look after I was drawn back to an issue I thought was long buried. This circuit was dropping 4-5% of its traffic at well below advertised capacity. The delivery method was ethernet over ATM/AAL5 (RFC1483). The usual suspects were rounded up and shot (duplex and error counters, purple switches, the cleaning lady).
The next question was overhead for cell tax and encapsulation. 12% seems to be the allowance by our supplier. IIRC this is the same as was used 10 years ago for ATM VCs delivered directly to routers (no ethernet involved); nanog has a few dusty posts about this number. It seems this is no longer enough. Perhaps average packet sizes have changed since then if 12% was ever a reasonable number.
In the end our supplier ramped overhead to about 20% and the pipe has performed nicely since. For an ethernet service, I don't understand why they don't simply provision the VC (VBR) as large as the customer bearer can handle and let the end devices do the shaping.
Michael's numbers including AAL5 padding give this new figure some context:
Upstream: cell tax = 14.84% (18.09% AAL5)
Downstream: cell tax = 18.35% (23.43% AAL5)
Is anyone aware of a recent study of packet size distribution on Internet traffic? CAIDA have some information from the AIX in 1999, NLANR from '97, and Sprint do a semi-regular dump from their ipmon project.
http://www.caida.org/analysis/AIX/plen_hist
http://www.nlanr.net/NA/Learn/packetsizes.html
http://ipmon.sprint.com/packstat/packetoverview.phpFrom netflow stats on this link, I see a similar story with 64, 576 and 1536B being the points of most interest.
The percentage of 1536B packets (~25) I think is larger now than previously.
Has anyone had similar frustrations with PVC dimensioning?
Does anyone else still use ATM? :)