[LEDE-DEV] ar71xx: performance decrease with kernel 4.4 (might be due to the qdisc/codel changes)

[Kevin Darbyshire-Bryant]

On 16/05/2016 10:14, Felix Fietkau wrote:
> On 2016-05-16 10:48, Hannu Nyman wrote:
>> I already said to Felix yesterday that I felt that with kernel 4.4 my ar71xx 
>> WNDR3800 seemed somehow more sluggish. Now I tested the matter with "flent".
>>
>> And sadly, with kernel 4.4 my router's performance decreases significantly. 
>> With kernel 4.1 the router achieves about 20% higher download throughput than 
>> with the 4.4 build :-(
>>
>> I used "flent" to measure wired connection throughput with
>> - two LEDE builds: r241 with kernel 4.1 and r253 with kernel 4.4
>> - two separate speed limits for SQM simple fq_codel QoS: 85000/10000 kb/s 
>> that should leave some CPU power free in the router, and 110000/15000 that 
>> should fully utilise the router's CPU.
>> - otherwise identical settings, all measurements made inside 30 minutes so no 
>> changes in traffic conditions
>>
>> The achieved speeds were:
>>
>> r241 kernel 4.1 - 85/10: 79 Mb/s down, 8.1 Mb/s up
>> r253 kernel 4.4 - 85/10: 67 Mb/s down, 8.5 Mb/s up
>>
>> r241 kernel 4.1 - 110/15: 85 down, 10.5 up
>> r253 kernel 4.4 - 110/15: 70 down, 10.8 up
>>
>> (ping latency stays at 16-17 ms with all four tries)
>>
>> With both SQM speed limits, the kernel 4.1 build performs significantly better.
>>
>> This performance decrease might be due to the kernel version bump to 4.4 or 
>> the qdisc/codel changes made to the 4.4 patches a few days earlier.
>>
>> This chart sums it up:
>> https://www.dropbox.com/sh/89pntkzjxydnn4c/AAC4x9cScJERL9Wfxm4k43kma?dl=0&preview=Kernel41_44_comparison.png
>>
>> Full flent data (summary pics & rrul data files) for all four tries is 
>> available in:
>> https://www.dropbox.com/sh/89pntkzjxydnn4c/AAC4x9cScJERL9Wfxm4k43kma?dl=0
> I've also noticed a throughput decrease on other platforms, I don't have
> any hard data on that though. I've tried tracking down the source of
> this regression, but haven't gotten anywhere with that yet.
>
> Once we're done with more urgent stuff, I plan to do some more work on
> optimizing the network stack.
I've already expressed my concern that the fq_codel batch drop backport,
backports more than it need to
http://lists.infradead.org/pipermail/lede-dev/2016-May/000349.html

Now whether the backport extras are a significant resource drain or not
isn't something I know or are in a position to test.

Another opinion/idea/guess for the pot :-)

Kevin