Current trends in capacity planning and oversubscription

Jeff Kell jeff-kell at utc.edu
Sat Nov 13 20:11:24 UTC 2010


On 11/13/2010 12:37 PM, Valdis.Kletnieks at vt.edu wrote:
> On Fri, 12 Nov 2010 18:52:20 EST, Sean Donelan said:
>> So what is a reasonable network capacity for 1,000 students now and in 5 
>> years.
> Just as LHC people and a school are different, I'm willing to bet that bandwidth
> "requirements" per student are different based on the school and its policies,
> and that they're to a large extent self-fullfilling.

There are "a number" of factors at play, but looking at a given snapshot
of network activity, some of the relevant metrics are the number of
active hosts [count inside clients if you are a content "consumer",
count outside clients of you are a content "provider"], the number of
active connections, and compare that to the bandwidth consumption.  It
doesn't really matter how many students you have, it's how many are
online and active at a given time.

> The "requirement" at a small liberal arts school with a fascist network usage
> policy "we block all bittorrent and any protocols we don't understand (i.e.
> most of them), and no network access in the dorms", will be different from a
> large engineering school that says "We'll provide bandwidth so you can explore,
> experiment, and learn, and we'll let you know if you're named in a copyright
> complaint".

That is true to a degree in terms of your degree of network neutrality. 
We do block most P2P, so I am not so certain of the bandwidth
consumption it would take these days if left open.  I do, however, know
that other traffic has simply risen to take it's place.  Well over half
of our bandwidth (less P2P) is streaming media, and a growing chunk of
that is Netflix.  I could take 50Mbps and divide it among a thousand web
surfers, or a hundred Youtube viewers, or a few dozen Netflix folks, or
one of our Athletics Dragonfly video servers. 

You can size it for a given population and a given application mix, then
be hit by the Obama inauguration, the Michael Jackson funeral, the World
Cup playoffs, or just a good healthy Patch Tuesday or Apple MacOS update.

> So "reasonable" bandwidth ends up depending on what the network admin thinks
> "reasonable" use of the network is...

And then there are limits to your ability to predict the "reasonable"
demands :-)  Which is where things like 95th percentile numbers start to
come in handy.

Another interesting metric is looking at "per user" consumption over
time.  Just as the applications have certain fringe cases that consume
an inordinate of bandwidth, you also have a few fringe case uses that
will account for an inordinate amount of bandwidth.  If your goal is to
accomodate these folks equally with the rest of the population, your
bandwidth requirements will be MUCH higher to accomodate them.

So with all that said... for what it's worth...  for the last 24hours,
our (sustained) inbound bandwidth (peak/avg) was 317Mbps / 107Mbps,
inside concurrently active host counts 2017/986, concurrent connection
counts 22.6K/10.8K.  That is with 11K students, 1500 fac/staff, and the
numbers being considerably higher during the week.  The 95th percentile
number is ~250Mbps for a typical weekday (24h), but we do "cap" dorm
traffic during classroom hours.

To backup another point brought up in this thread, the demand has
roughly doubled annually the last 3 years. 

Jeff




More information about the NANOG mailing list