AW: [support] performance test on MIPv6

[Trinks, Marcus (K-EFFI)]

Hi Sebastian,
thanks a lot for your detail description. 

When I run my tests again this morning I 
could not expirienced the behaviour again.
There was only a very little decrease and 
data througput.

Seems that I was to tired and made some mistake
yesterday. I am sorry...

cheers

Marcus

-----Ursprüngliche Nachricht-----
Von: support-bounces at l2tp.nautilus6.org [mailto:support-bounces at l2tp.nautilus6.org] Im Auftrag von Sebastien Decugis
Gesendet: Freitag, 22. August 2008 05:01
An: Support ML
Betreff: Re: [support] performance test on MIPv6

Hello,

This is an interesting measure, and also quite surprising. I have run a 
few performance tests last year with MIPv6 and IKEv2, and the loss of 
bandwith was not so important as far as I remember.

Could you please monitor the CPU usage of the components of your testbed 
(MNN, MR, HA, CN) while doing this test? I am thinking of a bug in iperf 
with recent kernels (> 2.6.21) for example.
Could you also capture some of the generated traffic in both situation 
(with and without MIPv6 running) to find out if fragmentation could be 
responsible for this?
Are you using IPsec to protect the payload between MR and HA?
I think it is also possible to retrieve some XFRM statistics that might 
be useful through the /proc interface, but I am not familiar with 
this... Maybe other people in this list can give more information?

Last point, the loss may be perfectly normal, depending on the 
performances of your HA and MR. The network path is different in your 
two tests, as follow:
Without MIPv6:   CN -------------------------------> MNN
With MIPv6: CN -------------> HA ====(tunnel)=====>MR-------------> MNN

In the first case, the HA and MR have only to forward the packets (I 
suppose you CN is attached outside the MNN network), but in the second 
case they have to encapsulate and decapsulate it, which is a far more 
computational job.
Also depending on where your CN and MR are located, you might be 
generating a lot of collisions in some ethernet subnets if for example 
CN is in the same subnet as MR.

I hope this information will help you to understand what causes the 
performance drop...

Also Romain, I believe there is a known performance issue with the 
tunnelctl that is used currently, but I don't remember if it will impact 
only handovers or all processing of packets...

Best regards,
Sebastien.

Trinks, Marcus (K-EFFI) a écrit :
>  
>
> -----Ursprüngliche Nachricht-----
> Von: Trinks, Marcus (K-EFFI) 
> Gesendet: Donnerstag, 21. August 2008 09:49
> An: 'support at ml.nautilus6.org.'
> Betreff: performance test on MIPv6
>
> Hello,
> I have got a question related to the network performance and MIPv6 implementation.
> Find a pdf of my testbed attached. 
>
> When I run iperf over my Ethernet Interface without running MIPv6, I reach a 
> throughput of about 90 Mbit/s. When I now start MIPv6 conducting the same test
> I reach a strongly fluctuating datarate between 0 Mbit/s and 2 Mbit/s. 
>
> Seems that MIPv6 influences the maximum throughput very strong. Can anybody of you
> please explain me that phenomenon?
>
> best regards
>
> Marcus
>   
> ------------------------------------------------------------------------
>
> _______________________________________________
> Support mailing list
> Support at ml.nautilus6.org
> http://ml.nautilus6.org/mailman/listinfo/support
>   
_______________________________________________
Support mailing list
Support at ml.nautilus6.org
http://ml.nautilus6.org/mailman/listinfo/support