Best Effort QoS Drop Profile Input

Mikael Abrahamsson swmike at swm.pp.se
Mon Aug 24 19:07:00 UTC 2009


On Mon, 24 Aug 2009, Brad Fleming wrote:

> We were never able to find a good best practice for configuring drop 
> profiles on edge devices. We've been using the following with OK results 
> but I was hoping to have some external, more experienced eyes take a 
> look...
>
> drop-profiles {
>   be_any {
>       fill-level 50 drop-probability 5;
>       fill-level 70 drop-probability 20;
>       fill-level 90 drop-probability 50;
>   }
>
> Has anyone ever run across a publicly documented best practice for drop 
> profile configuration? Does anyone have suggestions on ways to improve 
> the configuration above?
>
> Any input is much appreciated, thanks in advance!

I think it's fine, personally I've been known to configure core network 
WRED with 0% drop at 40ms and 100% drop at 60ms (reasoning is that 
basically no application will be well suited by having its packets delayed 
much more than 40ms), though at the lower speeds you have, your values are 
probably better. Hm, just realised your "fill level" is in percent, right? 
Then I'd stop dropping packets earler if I were you, having 0% drop up 
until 250ms is not really helping interactive applications, you probably 
want to induce drop earlier (at 40ms or so), so file transfers don't fill 
up the buffer but hopefully the fewer packets/s of an ssh session still 
has low probability of being dropped (and fast-retransmit can do it's job 
fairly unnoticable to the user).

So in your case, I'd have the first drop-probability much sooner, at 10% 
if you have 500ms buffering. Perhaps starting with a 1% at 10% and 3% at 
20% buffer fill rate.


-- 
Mikael Abrahamsson    email: swmike at swm.pp.se




More information about the NANOG mailing list