Re: [PATCH] CHOKe flow scheduler (0.8)

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 14 janvier 2011 à 04:58 +0100, Eric Dumazet a écrit :
> Le vendredi 14 janvier 2011 à 04:34 +0100, Eric Dumazet a écrit :
> 
> > Hmm, please wait a bit, I had another crash when I stopped my
> > bench/stress
> 
> I am not sure p->qavg is correctly computed.
> 
> Crash happened because choke_peek_random() was called while no packet
> was in queue.
> 
> With my params (min=10833 max=32500 burst=18055 limit=130000) this
> implies qavg was very big while qlen==0 !
> 
> qdisc choke 11: dev ifb0 parent 1:11 limit 130000b min 10833b max 32500b ewma 13 Plog 21 Scell_log 30
>  Sent 200857857 bytes 365183 pkt (dropped 1010937, overlimits 557577 requeues 0) 
>  rate 32253Kbit 7330pps backlog 17875996b 32505p requeues 0 
>   marked 0 early 557577 pdrop 0 other 0 matched 226680

Moving the qdisc_bstats_update(sch, skb); out of choke_enqueue() to
choke_dequeue(), I get nicer rate values (because packets that are
enqueued but CHOKed dont artificialy raise the packet/byte rates)

Now, rate properly matches my 10Mbit CBQ bandwidth :

qdisc choke 11: parent 1:11 limit 130000b min 10833b max 32500b ewma 13 Plog 21 Scell_log 30
 Sent 86470970 bytes 157418 pkt (dropped 127451, overlimits 48275 requeues 0) 
 rate 9947Kbit 2264pps backlog 17759368b 32288p requeues 0 
  marked 0 early 48275 pdrop 0 other 0 matched 39588


For other qdiscs, it is less easy because qdisc_bstats_update() call is
integrated in __qdisc_enqueue_tail() / qdisc_enqueue_tail(), so all
users shall be updated at once.