Thanks for your prompt response.
If it would be possible I want to know about the maximum throughput which you 
take from vpp in stateful mode (permit+reflect acl rule, session table) ?

Does vpp work scalable? I think the session table is bottleneck in stateful 
scenario and multi-threading? what 's your opinion?

Sent from Outlook<http://aka.ms/weboutlook>
________________________________
From: Andrew Yourtchenko <ayour...@gmail.com>
Sent: Wednesday, May 9, 2018 2:47 PM
To: Rubina Bianchi
Cc: vpp-dev@lists.fd.io
Subject: Re: [vpp-dev] VPP Scalability Problem

Dear Rubina,

You could take a look at “perf top” to see what could be going on. if you need 
help, let me know, I would be happy to look at it together.

Also, now as part of the work for 18.07 I am testing a couple of different 
approaches to change the processing for more performance, would you be 
interested to give it a try ?

--a

On 9 May 2018, at 13:19, Rubina Bianchi 
<r_bian...@outlook.com<mailto:r_bian...@outlook.com>> wrote:

Dear VPP Folks,
I have a problem in vpp scalability in statefull mode (by permit+reflect acl 
rules). I installed vpp on HPE ProLiant DL380 Gen9 Server in order to check the 
vpp performance and throughput scalability based on number of vpp worker 
threads.
Our appliance have 2 cpus that each cpu have 22 core (44 core in hyperthread 
mode). In our Scenario, 6 interfaces is considered. Each 2 interfaces is 
bridged and so we have 3 pair interfaces. Startup config file and other configs 
are added to attachment. With increasing the number of worker threads, I 
expected to see more throughput but it is not happened. I mean for example, 
with 18 workers (current config) I expect to see 45 Gbps (Because the best case 
throughput in sfr scenario is almost 15 Gbps per each pair interface) 
throughput (rx) on our DUT with 6 interfaces but the maximum throughput which 
is observed is approximately 24. And after that, increasing the number of 
worker threads not only did not have any effect on throughput but also in some 
cases decrease overall throughput catastrophically. Why?



Sent from Outlook<http://aka.ms/weboutlook>

<acls>
<startup.conf>
<vppctl>
<throughput.png>

Reply via email to