On 07 Jul 2004 14:57:48 -0400
jamal <hadi@xxxxxxxxxx> wrote:
> I seem to have hit the jackpot - all my emails to netdev are showing
> up and on time too.
>
> On Wed, 2004-07-07 at 14:10, Stephen Hemminger wrote:
> > Ok, I'll bite how would you do:
> >
> > Rate limit packet egress on a ethernet device (eth0) so it looks like a
> > slow DSL link (25 Kbps)
> > by not dropping packets but by pacing the data.
>
> Doesnt TBF work?
> rate 25kbit burst 90k should probably do it. Maybe i misunderstood the
> question.
TBF works but since the sender (on the same local machine) may go over it's
allocation,
it will drop packets.
For example, if I use tbf to simulate a slow 33k bits/sec link then TCP test
never
completes, it just hangs! TBF does work for intermediate sizes.
But if I use the pacing simulation it works.
>
> You may be able to avoid dropping but dont think you can guarantee it
> simply because you have finite buffers. At some point you will congest
> that queue and packets will be dropped; and if you dont limit your queue
> buffer size, sooner than later you are bound to hog all the system
> memory.
I understand that, every queue has to have a limit.
> Having said that, i have never seen a good arguement for why pacing
> traffic vs dropping to initiate a slowdown is better than the other.
> So in that case, a policer/meter should suffice.
|