Re: [RFC] sched: CHOKe packet scheduler (v0.2)

[Stephen Hemminger <shemminger@vyatta.com>]

On Fri, 7 Jan 2011 13:39:26 +0800
Changli Gao <xiaosuo@gmail.com> wrote:

> On Fri, Jan 7, 2011 at 12:55 PM, Stephen Hemminger
> <shemminger@vyatta.com> wrote:
> > On Thu, 06 Jan 2011 05:07:30 +0100
> > Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> >> Le mercredi 05 janvier 2011 à 11:21 -0800, Stephen Hemminger a écrit :
> >> > This implements the CHOKe packet scheduler based on the existing
> >> > Linux RED scheduler based on the algorithm described in the paper.
> >> >
> >> > The core idea is:
> >> >   For every packet arrival:
> >> >     Calculate Qave
> >> >     if (Qave < minth)
> >> >          Queue the new packet
> >> >     else
> >> >          Select randomly a packet from the queue
> >> >          if (both packets from same flow)
> >> >          then Drop both the packets
> >> >          else if (Qave > maxth)
> >> >               Drop packet
> >> >          else
> >> >               Admit packet with probability p (same as RED)
> >> >
> >> > See also:
> >> >   Rong Pan, Balaji Prabhakar, Konstantinos Psounis, "CHOKe: a stateless active
> >> >    queue management scheme for approximating fair bandwidth allocation",
> >> >   Proceeding of INFOCOM'2000, March 2000.
> >> >
> >> > Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
> >> >
> >>
> >> To be really useful in a wide range of environments, I believe that :
> >>
> >> - CHOKe should be able to use an external flow classifier (like say...
> >> SFQ) to compute a token and compare two skbs by this token instead of
> >> custom rxhash or whatever. (rxhash can be the default in absence of flow
> >> classifier). Probably you need to store the token in skb->cb[] to avoid
> >> calling tc_classify() several times for a given packet.
> >>
> >> http://lwn.net/Articles/236200/
> >> http://kerneltrap.org/mailarchive/linux-netdev/2008/1/31/667679
> >
> > Probably should split SFQ flow hash stuff into core code for reuse.
> >
> >
> 
> We need not do that, since we have sch_drr and cls_flow. :)

I prefer that the qdisc be useable without any explicit flow classification.
I.e like SFQ it should fall back to a sensible flow matching. DRR and others
put everything in one flow, if no filters are used.