Re: [aqm] AQM schemes: Queue length vs. delay based

[Naeem Khademi <[e-mail]>]

On Thu, Nov 14, 2013 at 12:33 AM, Anoop Ghanwani <[e-mail]>wrote:

>
>
>
> On Wed, Nov 13, 2013 at 3:23 PM, Curtis Villamizar <[e-mail]>wrote:
>
>>
>> Including unrealistic scenarios, like going from near zero traffic to 10
>> interfaces feeding one at full speed until overflow occurs, is
>> counterproductive.
>
>
> It is actually a problem that keeps many people busy because a number of
> data center switches have very high port count with very small buffers.
>  Some people address these buy using switches with bigger buffers, but
> that's not a luxury that everyone indulges in.
>

Fair point about the incast problem and I personally think AQMs designed
for data centers should address that.


> That is why I specifically asked the question at the AQM meeting about
> applicability of AQM to all types of networks/switches.  I was told the
> answer is "yes" and so I would like to see this scenario addressed as well.
>
>

Below is my personal opinion, but hopefully Fred can clarify this better
based on the AQM recommendations draft:

"applicability of AQM to all types of networks/switches" => "yes"

"applicability of *any* AQM to all types of networks/switches" => "no"


> Perhaps AQM cannot help this, but hopefully it won't hurt.  Trying to do
> fancy things with small buffers is challenging.
>

AQM will most likely to help from data centers to the access links, and so
on. But we may possibly need different AQMs for different network
scenarios; The fact that an AQMs should be auto-tunable doesn't imply that
it can be applied everywhere and we may need different auto-tunable AQMs
specifically designed for different networks (ideally better if we could
use fewer of them and they could work everywhere). I hope I'm not wrong.


>
> Anoop
>

Naeem