[ovs-discuss] Too many missed flows and high cpu usage

Tom van Leeuwen tom.van.leeuwen at saasplaza.com
Tue Oct 1 08:42:27 UTC 2013


Hi Ananth,

I'm wondering if your upgrade would also fix my issues on Citrix 
XenServer 6.1 
(http://openvswitch.org/pipermail/discuss/2013-September/011324.html). 
It was already pointed that there have been lots of performance 
improvements since version 1.4 and that I won't likely see my problems 
in version 1.11.

My Citrix guy told me that although it's technically possible to upgrade 
OVS in a Dom0, it's not supported. Any changes to Dom0 are not supported.
When you run into issues, which I've been doing a lot, the first thing 
they'll say: Install all hotfixes and bring back OVS to the supported 
stack. I wouldn't be surprised if they'd say: you've manually changed 
your Dom0, so now you must reinstall from scratch before we can help you.

For testing purposes I'm curious how you are going to upgrade OVS 
though, please keep me/us informed! I wish you luck.

Regards,
Tom

On 10/01/2013 09:37 AM, ananthan wrote:
> Thanks Justin,
>                       For Xenserver i am planning to upgrade from 1.4 
> to 1.9.Does 1.9.3 have wild carding feature.
> Regards,
> Ananth
>
>
> On Tue, Oct 1, 2013 at 11:22 AM, Justin Pettit <jpettit at nicira.com 
> <mailto:jpettit at nicira.com>> wrote:
>
>     Correct.  Multi-threading is going to be part of the 2.0 release.
>      The improvement you're seeing is from our adding support for
>     wildcarding in the kernel.  We've been calling it megaflows.  (As
>     opposed to the previous exact-match microflows that were
>     installed.)  In the new model, ovs-vswitchd tries to wildcard as
>     many fields as it can based on the configuration.  Since you are
>     just doing normal switching that matches on L2 addresses, most of
>     the fields are now wildcarded and these megaflows are covering the
>     bulk of your traffic.
>
>     --Justin
>
>
>     On Sep 30, 2013, at 10:18 PM, ananthan <ananthannair935 at gmail.com
>     <mailto:ananthannair935 at gmail.com>> wrote:
>
>     > Hi,
>     >      Tested ovs 1.11 on ubuntu 12.04 and recreated above
>     workload,this time i am so shocked to see that Load Average didnt
>     go more than 0.1%  which is nearly 80 times improvement when
>     compared to previous situation :) .It solves all the issue that we
>     were facing with old release,thanks for fixing all those
>     issue.Also i noticed that missed count is also far less.
>     >
>     > Waiting to upgrade it on xenserver 6.0.2
>     >
>     > but i dont see ovs-vswitchd threads using much cpu,so i assume
>     its not the multi-threading that fixed this,can you please tell me
>     the most important optimization that made to fix above things.
>     >
>     > Regards,
>     > Ananthan
>     >
>     >
>     >
>     >
>     > On Tue, May 28, 2013 at 9:12 PM, Justin Pettit
>     <jpettit at nicira.com <mailto:jpettit at nicira.com>> wrote:
>     >
>     > On May 28, 2013, at 8:39 AM, ananthan <ananthannair935 at gmail.com
>     <mailto:ananthannair935 at gmail.com>> wrote:
>     >
>     > > Hi,
>     > >      this is the only output for
>     > >
>     > > ovs-ofctl dump-flows xapi3
>     > >
>     > > duration=5120780.012s, table=0, n_packets=48344340859,
>     n_bytes=12069667659298, priority=0 actions=NORMAL
>     > >
>     > >     when i discussed this question on IRC someone pointed about :
>     > >
>     > > above default flow makes  ovs a standard learning/forwarding
>     L2 switch.
>     > >
>     > > Does this point has any part in high cpu usage?
>     >
>     > No.
>     >
>     > --Justin
>     >
>     >
>     >
>
>




More information about the discuss mailing list