[Click] ToDevice Cpu Load 99%

[Nicola Arnoldi]

May the problem I posted quite time ago be related to this notifier  
issue?

I use a scheduling, which is closely related to rrsched. Its  
notifiers are the same of rrsched.

The scheme is as follows:

...->class :: Classifier()
class[0]->NotifierQueue(x)->[0]scheduler;
class[1]->NotifierQueue(x)->[1]scheduler;
class[2]->NotifierQueue(x)->[2]scheduler;
class[3]->NotifierQueue(x)->[3]scheduler;

prio_sched :: PrioSched()
scheduler->[1]prio_sched
....->[0]prio_sched

where x is the length of NotifierQueue.

I tried to load the Classifier with a very high bitrate UDP traffic  
which is classified by the classifier and distributed to the four  
queues.
This is in a 802.11g system, with raw throughput of 26Mbps.
The scheduler is a weighted priority scheduler.
I noticed that, with UDP traffic, and queues shorter than 1000 pkts,  
the throughput is not distributed according to the weights.
Notifierqueue has got a sleepiness mechanism which avoids pulling  
from them when they hover around 1 or 2 pkts.

However the scheduler, which uses empty queues notifiers in order to  
learn if a queue is empty or not, skips pulling for queues  
(especially high prio ones) many times because of empty notifiers.
This happens only when queues are shorter than 600-700 pkts. This  
shouldn't be, as the packet rate injected into the queues is very high.

I also tried changing the sleepiness threshold, but it seems not  
affecting the issue.
I tried the configuration with short queues and RatedSource and  
RatedUnqueue in order to test it offline. It worked!
It seems that when applying it to a real scenario the whole thing  
collapses...

This fact is critical, because I have to mantaing very long queues  
which cause very high delays in my router!


Help needed!




Il giorno 26/gen/07, alle ore 22:04, Eddie Kohler ha scritto:

> Paolo,
>
> Do the recent notification fixes address this problem?
>
> Eddie
>
>
> Paolo Mailing wrote:
>> Hi to all,
>> I've a problem with the element ToDevice in linuxmodule,
>> in some complicated configuration the cpu load go fast to
>> 99%. I've some Queue plugged in a PrioSched and then
>> in another Priosched and then in ToDevice.
>> I've tried some configuration but I didn't understand the
>> problem.
>> The configuration is like this :
>>
>> FromDevice(eth0) ->
>> paintswitch :: PaintSwitch();
>> out_scheduler :: SimplePrioSched();
>> paintswitch[0] -> Queue -> [0] out_scheduler;
>> paintswitch[1] -> Queue -> [1] out_scheduler;
>> paintswitch[2] -> Queue -> [2] out_scheduler;
>> paintswitch[3] -> Queue -> [3] out_scheduler;
>> paintswitch[4] -> Queue -> [4] out_scheduler;
>> paintswitch[5] -> Queue -> [5] out_scheduler;
>> out_scheduler -> ToDevice(eth0);
>>
>> Thanks
>> Paolo
>> _______________________________________________
>> click mailing list
>> click at amsterdam.lcs.mit.edu
>> https://amsterdam.lcs.mit.edu/mailman/listinfo/click
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click