[altq 1468] Re: Another source of HFSC-ALTQ unfairness

[Oleg Cherevko <olwi@aq.ml.com.ua>]

On Mon, 29 Apr 2002, Kenjiro Cho wrote:

> One point.  In the case of link sharing, we should not remember the
> past too long.  Otherwise, a class which has used excess bandwidth
> in the past gets penalized badly. 

> So, HFSC uses consistent service
> curves only in the same parent backlog period,
^^^
To my understanding, no. This is not part of the original HFSC design.

> and resets vt of a
> newly backlogged class to the average of the min and the max of the
> already backlogged siblings. 
^^^
Yes, this is *the* mechanism that provides fairness of link sharing by
forgetting past (vt values) when appropriate.

> It's my understanding that the current
> link-sharing is designed to forget the past when the parent becomes
> idle.

BTW, what CMU folks do in their code (or rather how I understand what they
intended to do, for it seems to me that there are bugs in the code) is
that in the beginning of new parent backlog period (i.e. when first of 
its children becomes active) they traverse the list of all (inactive)
children of this parent and translate each virtual service curve to the
beginning of this new period by performing rtsc_min(). This is rather
costy. I believe that I know how to reduce this cost without compromising 
the idea. As I said earlier, I'd inform you on the results.

-- 
Olwi