Re: 10GE Click performance

Eddie Kohler <[email protected]>
Newsgroups gmane.network.routing.click
Organization Harvard University
Message-ID <[email protected]>
This is cool! I've added it to the click wiki under "publications"; 
please feel free to get a wiki login and clean up the cite.

Eddie


On 1/18/12 8:38 AM, Ricard Vilalta wrote:
> Hi Roman,
>
> I have recently published the following paper using click and 10GE
> transceivers.
> http://www.cttc.es/resources/doc/110728-hpsr-mplstp-final-46195.pdf
>
> I hope an extended version will be published soon in a journal.
>
> Best Regards,
> Ricard
>
> On 01/18/2012 02:29 PM, Luigi Rizzo wrote:
>> On Wed, Jan 18, 2012 at 2:13 AM, rchertov<[email protected]>   wrote:
>>> I finally got my hands on some 10GE equipment and started to play
>>> around with Click.  So I noticed the following.  On 2.6.24.7 patched
>>> kernel using 3.7.17 ixgbe driver, I get around 200K pps when running a
>>> node as a bridge (click pulled from git today).  However, when on
>>> exactly the same node I run the exactly same test but I use
>>> click-1.7.0rc1, then I can easily achieve 300K pps.  I have also tried
>>> 2.6.35.14-106.fc14.i686 using the latest Click and I still got around
>>> 200K pps.
>>>
>>>
>>> I am curious of people's experiences when running Click on 10GE
>>> equipment.  Has anybody got the RouteBricks MQ code running?  Everything
>>> compiled for me, but I get pretty strange packet forwarding performance,
>>> where the data is either delayed by quite a bit or it is just corrupted.
>>>
>>> My fancy one way bridge config
>>>
>>> fd :: FromDevice(eth2, PROMISC true, BURST 32)
>>>      ->   ctr1 :: AverageCounter
>>>      ->   q :: Queue(4096)
>>>      ->   ctr2 :: AverageCounter
>>>      ->   ToDevice(eth3, BURST 64);
>> remember that fetching timestamps, even staying in-kernel,
>> is extremely expensive (in the order of 250-500ns) so if you use
>> one of those elements in your pipeline you won't be able
>> to get decent performance.
>>
>> try to remove one or both counters and see if that improves
>> the throughput (this said, 200 or 300kpps really seems too
>> low to be explained by timestamps)
>>
>> cheers
>> luigi
>>
>>
>>
>
>
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.