Hi Tom,

Inline.

Cheers,
Pablo.
From: Tom Herbert <[email protected]>
Date: Sunday, 18 March 2018 at 16:28
To: "Pablo Camarillo (pcamaril)" <[email protected]>
Cc: dmm <[email protected]>, Uma Chunduri <[email protected]>
Subject: Re: [DMM] some test results of different network overlay methods


On Sun, Mar 18, 2018, 1:40 PM Pablo Camarillo (pcamaril) 
<[email protected]<mailto:[email protected]>> wrote:

Hi Tom, Kalyani,



If the SR processing of the two SIDs is done by the end host, then I believe 
the results for SR are obviously not valid.

Pablo,

Why isn't this valid?

What is the use case of an end linux host performing decapsulation or End SID 
processing?  By the time a packet reaches a Linux host terminating TCP it will 
contain only a single IP destination address == the host terminating the TCP 
session.

I expect that the only valuable use case to measure is encapsulation or 
insertion, see the  performance results 
https://irtf.org/anrw/2017/anrw17-final3.pdf for insert and encap on Linux.

If the end host is processing two SIDs, it means that you are doing twice SR 
processing. You will have an extra ‘End’ function processing, which involves 
updating the DA to the next segment, doing an IP lookup and forwarding.
Then you will restart and do again the processing. Overall, you do twice SR 
processing.

For example if I send  10 sids to a destination, all of which terminate on that 
destination, it will have to do SR END function processing 10 times.  This is 
true but there is no practical reason to do this.

In practice when host A sends a packet with 2 SIDs, one will be to host B and 
the next to host C.  Only B will do END function processing, C will simply 
receive the packet.  Therefore while your test is accurate it has no practical 
use.






Also in all the hardware-based routing platforms for which we have done interop 
https://tools.ietf.org/html/draft-filsfils-spring-srv6-interop-00#section-2, we 
have SRv6 running at linecard rate.



Note that I find this performance testing work interesting, and I’m willing to 
help you carry them. However, for next releases I would also consider 
leveraging projects like linux foundation fdio CSIT 
https://wiki.fd.io/view/CSIT/Description, in order to have a more methodical 
and accurate results (ILA has higher throughput than native IPv6 with a lower 
TPS).


To be clear, this is testing a real TCP stack and host terminating 
encapsulation (not just router performance). For this purpose the tests are 
methodical and accurate.
See above

The big difference in performance have a lot to do how well NICs deal with 
different encapsulations and EH in case of SR. The most interesting result, I 
think, is IPIP (and SR/IPIP) with drop in TPS. This indicates that device (in 
this case ixgbe) isn't getting 5-tuple hash for RSS. It's not parsing over 
headers to find ports for hash. Interestingly, it was able to parse into L4 
when just SR header is present. Capabilities for things like this will vary 
between NICs.

Indeed, this is very interesting. Thanks for pointing it out. I believe NICs 
providers will eventually fix this.

Tom




Thanks.



Cheers,

Pablo.

________________________________
De:Tom Herbert <[email protected]<mailto:[email protected]>>
Enviado:viernes, 16 de marzo de 2018 7:58 p. m.
Para:Uma Chunduri
Cc: dmm
Asunto: Re: [DMM] some test results of different network overlay methods

On Fri, Mar 16, 2018 at 12:40 PM, Uma Chunduri 
<[email protected]<mailto:[email protected]>> wrote:
> Great work, Thank you Kalyani & Tom.
>
> 2 quick questions:
>
> 1.  I presume SR inline is just SRH with 2 SIDs as mentioned - didn't see the 
> topology used. Do intermediate nodes handle these SIDs, with pointer update 
> in SRH?

Two hosts connected back to back. SR processing done by end host.

> 2.  Also for Geneve  - it's IP4 encap and VNI no TLVs?
>
No TLVs. GUE uses RCO extension. Other than that all the variants
should be out of the box with no options set. Encapsulations are
v4/v4, SRv6 and ILA are all IPv6.

I'll post the all the configuration scripts to github once I have some time.

Tom


> --
> Uma C.
>
>
> -----Original Message-----
> From: dmm [mailto:[email protected]] On Behalf Of Bogineni, Kalyani
> Sent: Friday, March 16, 2018 11:16 AM
> To: dmm <[email protected]<mailto:[email protected]>>
> Subject: [DMM] some test results of different network overlay methods
>
>
> All:
>
> Here are some raw performance test results based on our understanding of the 
> different network overlay methods. We welcome discussion and comments.
>
> Kalyani and Tom
> _______________________________________________
> dmm mailing list
> [email protected]<mailto:[email protected]>
> https://www.ietf.org/mailman/listinfo/dmm

_______________________________________________
dmm mailing list
[email protected]<mailto:[email protected]>
https://www.ietf.org/mailman/listinfo/dmm
_______________________________________________
dmm mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/dmm

Reply via email to