05:30:15 <anil_rao> #startmeeting taas
05:30:16 <openstack> Meeting started Wed Jul 27 05:30:15 2016 UTC and is due to finish in 60 minutes.  The chair is anil_rao. Information about MeetBot at http://wiki.debian.org/MeetBot.
05:30:17 <openstack> Useful Commands: #action #agreed #help #info #idea #link #topic #startvote.
05:30:19 <openstack> The meeting name has been set to 'taas'
05:30:24 <yamamot__> hi
05:30:26 <anil_rao> Hello
05:30:28 <kaz> hi
05:30:29 <soichi> hi
05:30:57 <anil_rao> #topic Performance measurement
05:31:12 <kaz> link: http://lists.openstack.org/pipermail/openstack-dev/attachments/20160727/04c7bcbc/attachment-0001.pdf
05:31:32 <anil_rao> Thanks Kaz for providing the new results.
05:31:38 <kaz> There is a mistake in the graph I showed last week IRC meeting.
05:31:49 <kaz> The attached file is a collected edition.
05:32:40 <anil_rao> I was reading on a blog that OVS is able to do around 260K pps
05:33:00 <anil_rao> Apparently this is a a new version with lots of perf enhancements.
05:33:39 <soichi> could you share URL of the blog?
05:33:54 <anil_rao> sure. give me a minute
05:34:19 <anil_rao> #link http://blog.ipspace.net/2014/11/open-vswitch-performance-revisited.html
05:34:26 <yamamot__> the vhost-net thread was sharing the cpu with something cpu-consuming?
05:35:06 <soichi> anil_rao: thiank you
05:36:55 <kaz> CPU pin down was enabled
05:37:48 <yamamot__> kaz: so the cpu was dedicated to the vhost-net instance?
05:39:56 <kaz> yamamot__: vhost-net is running on host OS, so it was not dedicated, I think.
05:41:37 <yamamot__> you mean vcpu threads are pinned?
05:42:01 <kaz> yamamot__: yes
05:42:42 <yamamot__> do you have a record of which cpu was running which thread?
05:43:26 <kaz> No I do not.
05:44:51 <anil_rao> If my understanding of the graph is correct it appears that when mirroring is not enabled, iperf is able to drive 95K pps.
05:45:35 <anil_rao> When mirroring is enabled you get 70K pps of production traffic + 70K pps of mirrored traffic.
05:45:50 <kaz> anil_rao: yes
05:46:28 <anil_rao> so when mirroring if off the 95K limit is either from iperf of the VMs themselves and not OVS. is this correct?
05:46:46 <yamamot__> the screenshot in p4 is for which case?
05:47:59 <kaz> yamamot__: this is for mirror is enabled case.
05:48:44 <kaz> anil_raw: yes, I guess
05:49:01 <anil_rao> I am curious to know why the non mirrored case tops off at 95K pps. Was vhost-net consuming 100% cpu in that case too.
05:51:38 <anil_rao> I was logged out of the IRC chat a few min back. Did the blog link i sent out reach you all?
05:52:10 <kaz> we know that vhost-net cosume 100% cpu , but we do not know the reason.
05:52:27 <soichi> anil_rao: okay
05:52:40 <anil_rao> soichi: Thanks
05:54:27 <anil_rao> This is an interesting result. I will need some time to fully understand the behavior.
05:54:52 <soichi> we will try to investigate
05:55:05 <anil_rao> I mean, when mirroring is enabled we are essentially seeing 140k pps (70K prod + 70K mirrored).
05:55:23 <soichi> yes
05:56:08 <anil_rao> So it seems to indicate that the 95K non mirrored limit is perhaps a limit with the dest VM not being able to receive faster than that rate.
05:56:49 <reedip> Hi sorry, I forgot that today we had a meeting
05:57:01 <anil_rao> Hi Reedip
05:57:06 <soichi> reedip: hi
05:57:11 <kaz> hi
05:57:31 <reedip> going through the meeting logs...
05:58:40 <anil_rao> Our lab is down because of an AC upgrade. So I will only have result for you all next week.
05:59:48 <anil_rao> any more on this topic, otherwise we can move to the next one on the agenda.
06:00:11 <muawiakhan> hi
06:00:15 <kaz> I guess the limit of 95K is a limit of source VM's CPU bound of vhost-net
06:02:39 <soichi> we will make additional experiments and report the result
06:02:53 <anil_rao> soichi: Thanks
06:03:00 <soichi> let's go no next topic
06:03:13 <anil_rao> #topic Design of traffic isolation by using flow based tunneling (cont.)
06:04:00 <soichi> Kaz found that Netron code was updated
06:04:28 <anil_rao> soichi: In what manner?
06:05:13 <kaz> I found that it is not able to use ":taas" in a neutron database because a new version of neutron has a strict type check of IP address.
06:06:38 <kaz> So we need to consider a new method of discovering IP address.
06:07:02 <soichi> yes, we need to revise our proposal
06:08:47 <soichi> we will submit revised one when we find a good idea.
06:09:05 <yamamot__> what's wrong with adding a column?
06:09:54 <soichi> yamamot__: it shoud be an idea, tahnk you
06:11:16 <soichi> move to next topic?
06:11:53 <anil_rao> soichi,kaz: I was wondering if we can return the right IP when the destination replies to the broadcast packet(s).
06:13:17 <kaz> broadcast packets need to send to appropriate tunnel.
06:15:04 <soichi> it means send to tunnels for production trafiic
06:15:14 <anil_rao> for taas we will always send via TAAS_SEND_FLOOD table, right?
06:16:23 <anil_rao> soichi,kaz: I will study this some more and we can discuss options in a later meeting.
06:16:54 <kaz> before mac learning, we send via TAAS_SEND_FLOOD table
06:17:45 <kaz> after mac lerning we will send via TAAS_SEND_UCAST table
06:19:29 <anil_rao> kaz: In today's exercise we are not really performing mac learning. Instead, we just learn about the location of the destination of a tap service instance. I am assuming that the same will be true with your flow based tunneling.
06:21:11 <kaz> thank you for your comment. I would like to discuss next week.
06:21:40 <anil_rao> sure. Let's move to the next topic.
06:21:49 <anil_rao> #topic Can not capture the packet, if the two VM on the same host
06:22:09 <anil_rao> #link Media:ingress_packet-20160720-r1.png
06:24:04 <anil_rao> Thanks for the detailed description of the problem. I am examining this issue...looks like the port vlan is not visible to the OVS flow table.
06:25:28 <anil_rao> I have reproduced the issue on my setup. I'll report back when I have a better handle on this.
06:26:02 <kaz> anil_rao: thank you
06:26:43 <anil_rao> What I did notice is that traffic coming from outside the host gets properly tagged by the flows in OVS.
06:27:33 <anil_rao> so this only affects ingress monitoring when the source is on the same host.
06:28:11 <kaz> anil_rao: that's right
06:28:40 <anil_rao> #topic  Open Discussion
06:28:48 <soichi> Let's vote to TaaS and related presentations.
06:28:56 <soichi> I found Vinay's presentation: "Tapping in NFV cloud : A real world showcase by Swisscom, Ericsson and IXIA." in the Telecom/NFV Operations track.
06:29:09 <anil_rao> +1
06:30:50 <anil_rao> we have run out of time. Next week then?
06:30:57 <anil_rao> #endmeeting