Re: Reasons for SFQ not being fair

["Покотиленко Костик" <casper@meteor.dp.ua>]

В Вто, 29/04/2008 в 15:53 +0300, Покотиленко Костик пишет:
> Hi there,
> 
> I've been using SFQ as the ending for every class of HTB for very  
> long. All working just as expected.
> 
> After implementing the abillity to do "limited speed, unlimited  
> traffic" clients I've experienced SFQ not being fair for those.  
> Precisilly, one of several connections gets bandwidth other stailed or  
> they run in turn.
> 
> Setup:
> 
> 1. There are 2 HTB classes for "limited speed, unlimited traffic"  
> clients each direction each ending with SFQ:
>   - base class (3:9900): for traffic < limit (2.5 Mbit/s for all)
>   - land class (3:9910): for landing traffic > limit (1 kbit/s for all)
> 
> 2. For each of those clients I insert 2 tc filters each direction:
>   - tc filter add dev $IN_DEV parent $IN_PARENT_CLASS protocol ip prio  
> 10 handle $HANDLE10 fw police rate $IN_RATE buffer 10k continue  
> classid $IN_BASE_CLASS
>   - tc filter add dev $IN_DEV parent $IN_PARENT_CLASS protocol ip prio  
> 11 handle $HANDLE10 fw classid $IN_LAND_CLASS
> 
> All this work perfect except SFQ fairness/flow speed division.
> 
> My thinkings:
> 
> # tc -s -d qdisc ls dev eth3 (filtered)
> qdisc sfq 9900: parent 3:9900 limit 128p quantum 1514b flows 128/1024  
> perturb 10sec
>   Sent 23753684098 bytes 71470965 pkt (dropped 0, overlimits 0 requeues 0)
>   rate 0bit 0pps backlog 0b 0p requeues 0
> qdisc sfq 9910: parent 3:9910 limit 128p quantum 1514b flows 128/1024  
> perturb 10sec
>   Sent 38036550 bytes 26648 pkt (dropped 1320085, overlimits 0 requeues 0)
>   rate 0bit 0pps backlog 162525b 126p requeues 0
> 
> In the "land" class there are too many droped packets. That's logical  
> because the rate is low (1 kbit/s) and there are only 128 packets  
> length queue. I think this is causing problems/unfairness because  
> droped packets means tcp retransmitions in best case and reconnects in  
> worst.
> 
> I didn't found a way to encrease the limit of 128 packet for the SFQ  
> it probably complile time limit.
> 
> Any hints/suggestions on how to tune SFQ?

Any hints? Maybe wrong list?

-- 
Покотиленко Костик <casper@meteor.dp.ua>