Ok thanks for the clarification.
I’ll check with CSIT and will see for the retry config when I get a chance to 
test this on my testbed.


From: "Damjan Marion (damarion)" <damar...@cisco.com>
Date: Tuesday, November 29, 2016 at 2:03 AM
To: "Alec Hothan (ahothan)" <ahot...@cisco.com>
Cc: vpp-dev <vpp-dev@lists.fd.io>
Subject: Re: [vpp-dev] native vhost-user perf improvements

Hi Alec,


On 29 Nov 2016, at 02:54, Alec Hothan (ahothan) 
<ahot...@cisco.com<mailto:ahot...@cisco.com>> wrote:

Hi Damjan,

Regarding the diffs, do you know in which error case(s) the retry will hit the 
most to cause such a big improvement in NDR?

It is not about error case. It is about cases where there is no enough space in 
the ring. At the point when we finished transmitting 1st incomplete batch of 
packets (incomplete because there was no enough free slots),
there is big chance the some space was already freed by the remote side. It is 
much cheaper to retry than to simply drop all work we did on the packets not 
transmitted in the 1st batch.

I have counted no less than 8 got done in the outer loop and it seems to me 
some of the error cases should not be retried at all (e.g. 

It is perfectly fine to set num_retry=0 in such cases.

Lastly, he max retry is hardcoded at 8 (line 1816), do you feel like this will 
work out the same retry value in all vpp ports?

I do'nt see why not, but fine tuning is possible. I was told that OVS is doing 
the same, same retry count (8).

If not it might be worth making it a vpp config parameter.

Contributions are welcome...

Also which performance test did you run to see the 40% improvement?

Maciek and Peter tested in their testbed.




vpp-dev mailing list

Reply via email to