[Cerowrt-devel] Latest build test - new sqm-scripts seem to work; "cake overhead 40" didn't

[Alan Jenkins]

On 19/06/2015, Sebastian Moeller <moeller0 at gmx.de> wrote:
> Hi Alan,
>
> excellent, thanks a million.
>
> On Jun 19, 2015, at 16:44 , Alan Jenkins
> <alan.christopher.jenkins at gmail.com> wrote:
>
>> Hi
>>
>> I guess I've done the complementary half to Seb's test :).  Basically
>> "cake overhead 40" didn't work, but that's the fault of cake in this
>> build.  Or tc, as Johnathan suggested.  (The "cake atm" part seems to
>> work, as per my previous test).
>
> 	Great!
>
>>
>> "tc qdisc" says "cake overhead 0", as Sebastian noticed.  And the test
>> results show "cake overhead 40" does not give a measurable
>> improvement.  But "tc stab overhead 40" does.
>>
>> I ran this test with the updated sqm-scripts and I think they're doing
>> the right thing.
>
> 	Thanks for testing this, especially as I can not due to a lack of an
> ADSL-link (and lack of cake actually, last I looked all I could find was
> cookies in my browser and a promise of pie in my router)
>
>>
>>
>> Method:
>>
>> I used the updated files from sqm-scripts,
>>
>> (once I remembered to mark them executable.  Lacking that causes a
>> failure with no error messages, because sqm-scripts checks before
>> running them :)
>>
>> but didn't bother updating & using luci-app-sqm.
>
> 	Ah, okay, I guess I did test this part with Dave’s help, so this should
> work with the most recent sqm.lua.
>
>>
>> The test was to compare netperf-runner results - ping during combined
>> upload & download - for "overhead 40" and "overhead 0".  I tested both
>> values of linklayer_adaptation_mechanism.
>>
>> I had to repeat 6 times (60s per run for each overhead) because of
>> random variation in the range of 3-4ms.  I alternated "overhead 40"
>> and "overhead 0" to try and exclude longer-term variation effects.
>>
>> With "stab overhead 40", median latency was better by about 3-4ms.
>> With "cake overhead 40", there is no such effect.
>
> 	Intersting, when I still had a 6M/1M ADSL link, I saw much larger latency
> under load increases when setting the per packet overhead to small, but I
> had my egress shaper running at 100% of line rate, so the system was rigged
> for maximum effect that way. How are your shapers typically set?

For this test I try to push it, today I used 95%.  I started trying
100%, which is still much better than unshaped.  I was scared off by
the random variation, I think it was higher at 100%.

For long term use I reduce it, because I've seen the line rate vary
slightly.  (1020k up today, 912 a while back.  Currently it reports a
"max" figure I don't understand, it's about 1100 despite being
rebooted daily.  16390k down).

Alan