17:17:31 <imaximets> #startmeeting ovn_community_development_meeting
17:17:32 <openstack> Meeting started Thu May  6 17:17:31 2021 UTC and is due to finish in 60 minutes.  The chair is imaximets. Information about MeetBot at http://wiki.debian.org/MeetBot.
17:17:33 <openstack> Useful Commands: #action #agreed #help #info #idea #link #topic #startvote.
17:17:36 <openstack> The meeting name has been set to 'ovn_community_development_meeting'
17:17:50 <blp> Thanks imaximets
17:17:58 <imaximets> I have some updates.
17:18:20 <imaximets> First is that OVSDB 2-Tier patch-set finally posted
17:18:51 <imaximets> #link https://patchwork.ozlabs.org/project/openvswitch/list/?series=241642
17:19:09 <imaximets> I also sent today a patch for raft
17:19:32 <imaximets> to transfer leadership before starting database compaction
17:19:48 <imaximets> #link https://patchwork.ozlabs.org/project/openvswitch/patch/20210506124731.3599531-1-i.maximets@ovn.org/
17:20:16 <imaximets> the  problem is that database compaction/snapshotting could take a lot of time if database is big
17:20:28 <blp> It's a good idea. Thanks for writing it up.
17:20:48 <imaximets> e.g 200-500MB of on-disk data.  It takes 2-15 seconds to create a snapshot
17:21:13 <imaximets> and cluster is not functional if the server that creates a snapshot is a leader.
17:21:35 <imaximets> transferring the leadership makes the system more stable.
17:21:58 <imaximets> and the last update:
17:21:59 <blp> Wow, that's a lot. I never did measurements.
17:23:00 <imaximets> I also worked few weeks ago on a patch for OVN to collapse ARP responder flows.
17:23:22 <blp> Excellent.
17:23:29 <imaximets> On some setups these flows could take half of the southbound db.
17:23:39 <blp> wow
17:23:50 <imaximets> In my case it was almost 250MB out of 500MB database.
17:24:10 <imaximets> And the change actually reduced the db size in half.
17:24:22 <numans> imaximets++
17:24:28 <numans> that's great
17:24:47 <imaximets> I have a patch, but I don't have a ddlog part yet.
17:25:21 <imaximets> it looks like this: https://github.com/igsilya/ovn/commit/732365025682cdb2987d601d3caedee1a94dfcf7
17:26:21 <imaximets> the idea is to simplify the actions (i.e. remove dependency from addresses) and put all load balancer IPs into one lflow instead of creating a separate lflow for all of them.
17:26:55 <imaximets> Alternatively this could be done with the address set.  I'll probably will go with the address set for the final patch.
17:27:03 <blp> Wonderful.
17:27:31 <imaximets> IIRC, tomorrow is the deadline to submit v1 for 20.06?
17:27:35 <imaximets> *21.06
17:28:33 <imaximets> So, I will post some version of the patch tomorrow.  And will work it out later if will not be able to figure out ddlog parts until then.
17:28:38 <zhouhan> imaximets combining flows may have some implication on I-P processing (maybe that's not a problem for this case, but we need to be careful)
17:29:38 <blp> I upgraded to Fedora 34 and therefore GCC 11 and I'm getting some new warnings for OVS. I didn't notice anyone posting patches for these already, has anyone else?
17:29:44 <imaximets> zhouhan, sure.  But in this case shrinking the database in half is kind of important too, because ovsdb-server just cant handle it.
17:30:15 <imaximets> blp, I have a 'may be uninitialized' in netdev-dummy.
17:30:37 <blp> OK, this is different. I'll fix and post.
17:30:42 <zhouhan> imaximets yes. CPU & MEM sometimes need trade-offs
17:31:01 <imaximets> blp, thanks!
17:31:21 <imaximets> ok.  that was it from my side.
17:31:28 <imaximets> who will go next?
17:31:51 <blp> I just have a quick report
17:32:01 <imaximets> blp, go ahead.
17:32:02 <blp> which is that I'm getting back to fixing up my patch series and will repost.
17:32:24 <blp> Sorry I disappeared for a while.
17:33:00 <_lore_> can I go next? quite fast
17:33:29 <imaximets> _lore_, sure, if that's all from blp?
17:33:38 <_lore_> blp: yes, sorry
17:33:54 <blp> Oh, I think that's all.
17:34:08 <blp> If I think of more, I'll add it.
17:34:20 <blp> Oh, one thing is that we got a paper about OVS accepted at SIGCOMM.
17:34:51 <zhouhan> congrats!
17:34:58 <imaximets> blp, cool!
17:35:21 <imaximets> ok, _lore_ , your turn.
17:35:29 <_lore_> ok, last week I worked on CoPP series, I added unit/system tests and remove per-port metering, posted v1 upstream
17:35:35 <_lore_> http://patchwork.ozlabs.org/project/ovn/list/?series=241400
17:36:04 <_lore_> I guess we can continue discussing from v1 about possible changes on API and so on
17:36:49 <_lore_> I added a fix for localport --> localnet traffic
17:36:50 <_lore_> http://patchwork.ozlabs.org/project/ovn/patch/8008fa9867d210cf18ad31f912535f2c14e85c43.1620151078.git.lorenzo.bianconi@redhat.com/
17:37:19 <_lore_> and I have a pending debug patch out for review
17:37:21 <_lore_> http://patchwork.ozlabs.org/project/ovn/patch/3a9ca2d94478a3b2db51d719233c9a28cafcf75b.1619268349.git.lorenzo.bianconi@redhat.com/
17:37:29 <_lore_> that's all from my side, thx
17:38:02 <zhouhan> I can go next
17:38:12 <imaximets> _lore_, thanks!
17:38:16 <imaximets> zhouhan, sure.
17:38:33 <zhouhan> This patch still needs review: https://patchwork.ozlabs.org/project/ovn/patch/20210504013332.1689887-1-hzhou@ovn.org/
17:38:50 <zhouhan> The last version addressed Mark M
17:39:04 <zhouhan> 's comments about test case improvement
17:39:24 <mmichelson> zhouhan, I'll have a look at that this afternoon. I plan to review that plus a patch of numans's
17:39:28 <mmichelson> (btw, I'm back now)
17:39:40 <zhouhan> mmichelson cool. thanks a lot!
17:40:34 <zhouhan> I was also checking DP group. I think there is a small problem on the MC_UNKNOWN flow handling. I have a fix and will send out today
17:41:24 <imaximets> zhouhan, we're not using DP groups for multicast groups.  Or the issue is that we are doing that somewhere?
17:42:02 <zhouhan> imaximets yes, I think it is just missed for the MC_Unknown group. It is using DPG :)
17:42:18 <zhouhan> the fix is simply not using DPG for the flow
17:42:39 <imaximets> zhouhan, oh, ok.  Sounds good. :)
17:43:01 <zhouhan> It also seems that DP group could have a higher CPU cost in ovn-controller when dealing with port group changes for ACL flows, because each change would have be bigger impact with the DPG enabled. But I will study more on this.
17:43:25 * imaximets thinks that we still need to fix dependencies somehow between multicast groups and lflows.
17:43:32 <zhouhan> imaximets also, any plan to make DPG work in DDlog version? (Or is it already supported but I missed somehow?)
17:44:03 <mmichelson> I thought it was supported there
17:44:09 <imaximets> zhouhan, should be supported.  There is a test for them.  blp should know, I think.
17:44:45 <mmichelson> When you run a test with OVN_FOR_EACH_NORTHD now, it will run four versions: C with DP groups, C without DP groups, DDLog with DP groups, and DDLog without DP groups
17:45:27 <imaximets> zhouhan, "ovn -- logical gatapath groups" should work for all implementations, so it should be supported.
17:46:20 <blp> datapath groups? I think that they are supported.
17:46:21 <blp> in ddlog
17:46:43 <zhouhan56> test
17:46:51 <blp> Wow, that's a lot of variations, 4x for each of those?
17:47:11 <mmichelson> Yep
17:47:37 <imaximets> blp, unit tests are a bit slow now. :)  but that's for good.
17:47:40 <blp> The worst part of the OVN_FOR_EACH_NORTHD stuff is that it doesn't report line numbers very well.
17:47:52 <blp> If anyone figures out how to do that better, that'd be great.
17:48:10 <blp> zhouhan56: we see you
17:48:43 <zhouhan> hi I'm back
17:48:52 <zhouhan> zhouhan56 cannot talk :(
17:49:01 <imaximets> zhouhan, hi. :)
17:49:12 <zhouhan> sorry I was disconnected for a while
17:49:32 <imaximets> zhouhan, we concluded that DPG supported in ddlog.
17:49:46 <blp> zhouhan: welcome back
17:49:49 <zhouhan> imaximets oh, great. Sorry I missed that
17:50:32 <imaximets> zhouhan, for the performance, IIRC, there was a lot of issues with I-P engine, so I had to sacrifice some performance on the ovn-controller side.
17:50:48 <imaximets> i.e. I-P is not ideal with DPG.
17:51:37 <zhouhan> imaximets: yes there is some challenge, but I will see if there is a way to make DPG work well with I-P
17:51:51 <imaximets> zhouhan, that would be great.  thanks!
17:51:51 <zhouhan> MEM and CPU are both important for some of our environment :)
17:52:03 <zhouhan> the last thing I said before disconnecting was: imaximets the dependency part I think I also have a fix in mind, will work on it as well
17:52:44 <zhouhan> Apart from these, I did review of numans CT patches
17:52:48 <imaximets> zhouhan, nice.  this also would be great to have.
17:53:23 <zhouhan> Thanks numans for the I-P improvement, but I'd like to discuss more on that patch
17:53:28 <zhouhan> that's my update
17:53:50 <imaximets> ok.  Who wants to go next?
17:53:59 <mmichelson> I can go next
17:54:13 <imaximets> mmichelson, all yours.
17:54:36 <mmichelson> I have two main things I've worked on recently. First is the ARP-related patch series. I posted a v7 of it where I ended up whitelisting a warning message from ovn-controller that had been causing systems tests to fail
17:55:00 <mmichelson> I've talked that particular issue over with imaximets and dceara and dceara suggested a possible workaround that I'm going to look into.
17:55:15 <mmichelson> v7 also triggered 0-day robot since the line length checks are working again in checkpatch.py
17:56:13 <mmichelson> The other thing I'm working on is an issue where the same conjunction() action can appear multiple times in a flow. I figured out the issue, it's a one-line fix in expression code. The explanation of the issue and the fix will take probably 50x more lines than the actual fix.
17:56:33 <mmichelson> That particular fix should hit the mailing list later today
17:57:01 <mmichelson> Other than that, we're approaching soft freeze tomorrow. So any patches that you  want included in 21.06.0 should be posted by tomorrow.
17:57:11 <mmichelson> That's all from me.
17:57:32 <blp> I posted my GCC 11 series: https://patchwork.ozlabs.org/project/openvswitch/list/?series=242511
17:57:40 <blp> (It's for OVS, not OVN.)
17:58:23 <imaximets> blp, thanks, I'll take a look.
17:58:27 <mmichelson> blp, is there a pending OVN patch, too?
17:58:56 <zhouhan> mmichelson shall we release 20.03.1 anytime soon?
17:59:06 <zhouhan> sorry, 21.03.1
17:59:43 <mmichelson> zhouhan, sure, we can do that when we release 21.06.0
18:00:07 <blp> mmichelson: I'll look at OVN now.
18:00:18 <zhouhan> mmichelson ok, thanks!
18:01:23 <imaximets> Good.  Anyone else has updates?
18:02:34 <imaximets> OK.  In this case we can close this meeting.  Have a nice day!
18:02:39 <imaximets> #endmeeting