Contention/Oversubscription maths

Adam Armstrong lists at
Fri May 27 13:07:49 UTC 2011

On 27/05/2011 14:02, Jacob Broussard wrote:
> I don't use almost any bandwidth outside of Netflix, Steam game 
> downloads, and getting my daily dose of streaming starcraft videos and 
> ntop tells me I averaged 1.7mbps over the last month.  Mind you this 
> is on an 8mbps peak connection.  With peak speeds of 8m I would be 
> pissed if I was getting 500k, much less if I had a 100m connection and 
> got less than a meg.  I have no doubt that if I had a faster 
> connection that I would have used even more bandwidth... With the 
> popularity of streaming video now a 1000:1 or even a 100:1 
> oversubscription rate is almost definitely just going to cause you 
> headaches...  Back in my days of noc, 9 out of 10 bandwidth AUP 
> abusers weren't even using torrents, they were almost all netflix or 
> people that got rid of cable to watch video online.
I suspect your average usage is an order of magnitude higher than that 
of the 'average' user.

This is one of the effects that makes it hard for me to imagine the 
traffic profile, as my own usage patterns (and those of my circle of 
friends) is not what one would call typical.

For example if 300KB/sec average peak holds true with only 1000 users, 
that's only 300Mbit usage, giving 700Mbit headroom. I suspect this will 
only become true with 10000 users on 10G though, and that 1G links will 
be much more bursty as a single customer can push 10% of the backhaul speed.

Remeber that 1000:1 contention *doesn't* mean that you only get 1000th 
of the backhaul speed, it means the *peak average* usage of all users 
should be under 1Mbit. Your peak is most likely 8Mbit, but there are 
probably 40 users who use very liuttle for every one of you, and they 
may only double the peak average to 16mbit, thereby giving a much lower 
peak average for all 41 of you (if you see what i mean?)

I suck at maths, and I'm pretty sure I was at home playing Tekken 2 when 
I should have been in statistics class :)


More information about the NANOG mailing list