[Codel] OpenWRT wrong adjustment of fq_codel defaults (Was: fq_codel_drop vs a udp flood)

[Roman Yeryomin]

On 6 May 2016 at 21:43, Roman Yeryomin <leroi.lists at gmail.com> wrote:
> On 6 May 2016 at 15:47, Jesper Dangaard Brouer <brouer at redhat.com> wrote:
>>
>> I've created a OpenWRT ticket[1] on this issue, as it seems that someone[2]
>> closed Felix'es OpenWRT email account (bad choice! emails bouncing).
>> Sounds like OpenWRT and the LEDE https://www.lede-project.org/ project
>> is in some kind of conflict.
>>
>> OpenWRT ticket [1] https://dev.openwrt.org/ticket/22349
>>
>> [2] http://thread.gmane.org/gmane.comp.embedded.openwrt.devel/40298/focus=40335
>
> OK, so, after porting the patch to 4.1 openwrt kernel and playing a
> bit with fq_codel limits I was able to get 420Mbps UDP like this:
> tc qdisc replace dev wlan0 parent :1 fq_codel flows 16 limit 256

Forgot to mention, I've reduced drop_batch_size down to 32

> This is certainly better than 30Mbps but still more than two times
> less than before (900).
> TCP also improved a little (550 to ~590).
>
> Felix, others, do you want to see the ported patch, maybe I did something wrong?
> Doesn't look like it will save ath10k from performance regression.
>
>>
>> On Fri, 6 May 2016 11:42:43 +0200
>> Jesper Dangaard Brouer <brouer at redhat.com> wrote:
>>
>>> Hi Felix,
>>>
>>> This is an important fix for OpenWRT, please read!
>>>
>>> OpenWRT changed the default fq_codel sch->limit from 10240 to 1024,
>>> without also adjusting q->flows_cnt.  Eric explains below that you must
>>> also adjust the buckets (q->flows_cnt) for this not to break. (Just
>>> adjust it to 128)
>>>
>>> Problematic OpenWRT commit in question:
>>>  http://git.openwrt.org/?p=openwrt.git;a=patch;h=12cd6578084e
>>>  12cd6578084e ("kernel: revert fq_codel quantum override to prevent it from causing too much cpu load with higher speed (#21326)")
>>>
>>>
>>> I also highly recommend you cherry-pick this very recent commit:
>>>  net-next: 9d18562a2278 ("fq_codel: add batch ability to fq_codel_drop()")
>>>  https://git.kernel.org/davem/net-next/c/9d18562a227
>>>
>>> This should fix very high CPU usage in-case fq_codel goes into drop mode.
>>> The problem is that drop mode was considered rare, and implementation
>>> wise it was chosen to be more expensive (to save cycles on normal mode).
>>> Unfortunately is it easy to trigger with an UDP flood. Drop mode is
>>> especially expensive for smaller devices, as it scans a 4K big array,
>>> thus 64 cache misses for small devices!
>>>
>>> The fix is to allow drop-mode to bulk-drop more packets when entering
>>> drop-mode (default 64 bulk drop).  That way we don't suddenly
>>> experience a significantly higher processing cost per packet, but
>>> instead can amortize this.
>>>
>>> To Eric, should we recommend OpenWRT to adjust default (max) 64 bulk
>>> drop, given we also recommend bucket size to be 128 ? (thus the amount
>>> of memory to scan is less, but their CPU is also much smaller).
>>>
>>> --Jesper
>>>
>>>
>>> On Thu, 05 May 2016 12:23:27 -0700 Eric Dumazet <eric.dumazet at gmail.com> wrote:
>>>
>>> > On Thu, 2016-05-05 at 19:25 +0300, Roman Yeryomin wrote:
>>> > > On 5 May 2016 at 19:12, Eric Dumazet <eric.dumazet at gmail.com> wrote:
>>> > > > On Thu, 2016-05-05 at 17:53 +0300, Roman Yeryomin wrote:
>>> > > >
>>> > > >>
>>> > > >> qdisc fq_codel 0: dev eth0 root refcnt 2 limit 1024p flows 1024
>>> > > >> quantum 1514 target 5.0ms interval 100.0ms ecn
>>> > > >>  Sent 12306 bytes 128 pkt (dropped 0, overlimits 0 requeues 0)
>>> > > >>  backlog 0b 0p requeues 0
>>> > > >>   maxpacket 0 drop_overlimit 0 new_flow_count 0 ecn_mark 0
>>> > > >>   new_flows_len 0 old_flows_len 0
>>> > > >
>>> > > >
>>> > > > Limit of 1024 packets and 1024 flows is not wise I think.
>>> > > >
>>> > > > (If all buckets are in use, each bucket has a virtual queue of 1 packet,
>>> > > > which is almost the same than having no queue at all)
>>> > > >
>>> > > > I suggest to have at least 8 packets per bucket, to let Codel have a
>>> > > > chance to trigger.
>>> > > >
>>> > > > So you could either reduce number of buckets to 128 (if memory is
>>> > > > tight), or increase limit to 8192.
>>> > >
>>> > > Will try, but what I've posted is default, I didn't change/configure that.
>>> >
>>> > fq_codel has a default of 10240 packets and 1024 buckets.
>>> >
>>> > http://lxr.free-electrons.com/source/net/sched/sch_fq_codel.c#L413
>>> >
>>> > If someone changed that in the linux variant you use, he probably should
>>> > explain the rationale.
>>
>> --
>> Best regards,
>>   Jesper Dangaard Brouer
>>   MSc.CS, Principal Kernel Engineer at Red Hat
>>   Author of http://www.iptv-analyzer.org
>>   LinkedIn: http://www.linkedin.com/in/brouer