[c-nsp] latency measurement equipment

[Mack McBride]

You may be able to tell your vendor you need a third party eval and have them pay for it.
Most vendors in the low latency market should not balk at that or have already done such an eval.
If they don't have one and they balk at the idea then buy from someone who has one already.

Mack

-----Original Message-----
From: cisco-nsp-bounces at puck.nether.net [mailto:cisco-nsp-bounces at puck.nether.net] On Behalf Of Jeff Bacon
Sent: Monday, June 06, 2011 1:48 PM
To: cisco-nsp at puck.nether.net
Subject: Re: [c-nsp] latency measurement equipment

> From: Mack McBride <mack.mcbride at viawest.com>
> 
> The only application where I have seen profitability in decreasing
> latency is the financial trading industry (see definition of
> profit).  

Even in the financial industry, I would say there is a log curve of
diminishing returns for reduced latency - the costs simply go
astronomical. That said, there are other industries that care too, at
least to a certain level. 

There are folks who are busily straightening the fiber path between New
York and Chicago (a friend of mine is involved in those projects), and
they aren't far from simply taking a 700 mile long strand of beryllium
monocrystal, attaching two large trucks to either end, and pulling. :) 

My interest is primarily to verify that the numbers in the glossies
really are realistic for the kit and for what we're doing with it, in
the skeptical "show me" sense - I want to believe, but I've had a number
of reasons lately to not believe (not just the FWSM) and I do need to
know. I'm sure I could spend a lot of time rigging up a way to do it
with commodity kit - say with passive taps going to a single central
linux box using PF_RING to cut out most of the kernel overhead - but if
I can get proper test equipment without blowing the bank then I'd rather
do that. 

> One other consideration is the end equipment.  If a
> standard linux/windows kernel is used then the push for decreased
> latency is wasted.

acknowledged. that's being dealt with as well. though I would say you
don't need a real-time kernel to get down to 10usec or so, just clever
playing with the scheduler. 

_______________________________________________
cisco-nsp mailing list  cisco-nsp at puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/