[88984] in North American Network Operators' Group
Re: SUMMARY - 95th percentile calculation
daemon@ATHENA.MIT.EDU (Bill Stewart)
Tue Feb 28 05:10:23 2006
Date: Tue, 28 Feb 2006 02:09:52 -0800
From: "Bill Stewart" <nonobvious@gmail.com>
To: nanog@merit.edu
In-Reply-To: <20060227194459.GA74516@svcolo.com>
Errors-To: owner-nanog@merit.edu
I'm not operationally involved at AT&T, but what I've been told is that we =
track
5-minute samples in both directions, and do the 95th% calculation on
all the samples,
as opposed to tracking 95% of inbound-only or outbound-only or
max(in,out) samples.
On 2/27/06, Jo Rhett <jrhett@svcolo.com> wrote:
>
> My thesis that I wanted to prove or disprove was that the market seems to
> have standardized how they calculate 95th percentile. Years ago if you
> asked all your uplinks you'd get wildly different answers. Averages or m=
ax
> values over days, weeks or months. Rarely would two answers jive.
>
> But it appeared that most providers are now using 5 minute averages as th=
e
> value input over a billing month for 95th% calculation. Was this truly
> standard now?
>
> The answer is yes. All of our paid transit links use 5 minute averages,
> and all but one of the replies to my query here did too.
>
> 8 providers use 5 minute averages of 30 second counter checks as the base
> value calculated over the billing month.
>
> 2 providers use 5 minute averages of 1 minute counter checks as the base
> value calculated over the billing month.
>
> 8 providers use 5 minute implicit averages (a single counter check) as th=
e
> base value calculated over the billing month.*
>
> 1 provider uses the maximum value found in 5 minutes of 1 minute counter
> checks for the 5 minute value, and calculates that over the billing month=
.
>
> 1 provider uses 5 minute implicit averages (a single counter check) but
> calculates that over a fixed 30 day month. (I didn't ask what they did wi=
th
> the extra days)
>
> At first glance my tendancy would be to say that 20 answers is hardly
> authoritative, but the 30 second counters (first listed) comprise 6 of th=
e
> top US bit movers, and 2 of the largest bit pushers so I'm satisfied with
> the results for my purposes.
>
> * One very large bit mover that I know uses 32 bit counters claimed to on=
ly
> be checking every 5 minutes. We're pushing enough to them to rotate that
> counter in that interval, so I actually disbelieve this answer but I'm
> reporting it as they said it.
>
> --
> Jo Rhett
> senior geek
> SVcolo : Silicon Valley Colocation
>
--
----
Thanks; Bill
Note that this isn't my regular email account - It's still experimental so =
far.
And Google probably logs and indexes everything you send it.
