Il 28/04/2013 13:45, Luca dariz ha scritto:
> Il 26/04/2013 22:00, Luca dariz ha scritto:
>> Il 26/04/2013 16:35, Conor O'Gorman ha scritto:
>>> On Fri, 2013-04-26 at 14:13 +0200, Luca dariz wrote:
>>>> Use a tasklet to handle incoming packets. Fix #12917.
>>>>
>>>> Incoming packets are now processes in a tasklet instead of in the
>>>> irq handler; this should improve latency.
>>>>
>>>> This patch is based on a previous version of ltq-atm driver, which
>>>> did implement a tasklet.
>>>>
>>>> It has been tested on a arv4518pw with a
>>>> Lantiq Danube for about a month and it seems to work well.
>>>
>>> And how much 'better' is it?
>>>
>>
>> I did't measure latency with this patch and without it, so i can't tell
>> exactly.
>>
>> Luca
> 
> A quick test with cyclictest
> (https://rt.wiki.kernel.org/index.php/Cyclictest):
> 
> without the patch:
> 
> cmdline: ./cyclictest  -n
> # /dev/cpu_dma_latency set to 0us
> policy: other/other: loadavg: 0.51 0.33 0.27 1/43 3241
> T: 0 ( 3241) P: 0 I:1000 C: 144288 Min: 40 Act: 93 Avg: 163 Max: 4456
> 
> with the patch:
> 
> cmdline: ./cyclictest -n
> # /dev/cpu_dma_latency set to 0us
> policy: other/other: loadavg: 0.78 0.44 0.22 1/44 4981
> T: 0 ( 3452) P: 0 I:1000 C:  82427 Min: 39 Act: 199 Avg: 252 Max: 18179
> 
> All latencies are measures in microseconds.
> 
> Luca

I forgot: during the test, a big download was in progress, to stress the
rx path a bit.

Luca
_______________________________________________
openwrt-devel mailing list
openwrt-devel@lists.openwrt.org
https://lists.openwrt.org/mailman/listinfo/openwrt-devel

Reply via email to