[c-nsp] Cisco and microbursts

[Saku Ytti]

On 19 May 2016 at 15:15, Adam Vitkovsky <Adam.Vitkovsky at gamma.co.uk> wrote:

> But one has to strike the balance right as having these too deep can introduce too much delay into the network.
> At some point you want the packets to get dropped so TCP can do its thing and adjust.

Alas there isn't any perfect solution. Either you allow high TCP rate
on high RTT towards lower rate port or you make sure there isn't
useless delay on actual congestion. You really can't have both.
This is why buffer depth should be configured to match the
goals/customer requirements. Very very often I see people allocating
all remaining buffers in BE class, regardless of hardware they are
running, implying they don't have any policy how long they'll buffer
BE. Everyone should decide what they want and configure box to match
that, at least then they'll know it, and if decision proves bad, they
can change the decision.

One thing I'd like to see, is ability to configure true temporal
buffers. Imagine that my connection is congested pretty much always,
but during office hours I'm low on BE and I'm loaning it's capacity
for AF, i.e. AF is out-of-contract but can loan capacity from BE. But
outside office hours, situation is reversed, BE is out-of-contract,
but can loan capacity from AF.
Now if I'm setting fixed byte size buffers, I have variable queueing
delay, as offered rate in AF and BE is very different at different
time of the day. What I'd really want to do, is to tell AF that it
won't buffer more than 5ms and BE not more than 15ms, regardless of
offered rate. So I'd know there is consistent buffering delay during
congestion.
Today I just kinda have to make compromise to buffer little bit too
little at one hour and little bit too much at another hour.


-- 
  ++ytti