buffer sizing - Re: [c-nsp] 12008 replacement, 7300?

[dave o'leary]

At 09:22 AM 5/13/2005 -0700, David Barak wrote:
>some comments in-line.
>
>--- Matthew Crocker <matthew at crocker.com> wrote:
>.....
> > If you are taking big pipes and feeding little pipes
> > you'll want the
> > bigger packet buffers.  high latency is better than
> > dropped packets.
>
>depends on your traffic type - there have been some
>talks at NANOG about how small buffers can actually
>improve TCP throughput, and how large buffers can
>decrease it.  Certainly large buffers are bad for
>jitter, so if you're doing circuit-em or other traffic
>where Jitter is an issue, you'll want really small
>buffers.

or alternatively, you need mechanisms for traffic engineering,
including classifying packets and scheduling them appropriately -
with a network designed such that no queue should have more
traffic with low latency requirements than the packets can
accept.
                                         dave