18:19:00 <dfarrell07> #startmeeting cperf
18:19:00 <collabot> Meeting started Thu Jul 21 18:19:00 2016 UTC.  The chair is dfarrell07. Information about MeetBot at http://wiki.debian.org/MeetBot.
18:19:00 <collabot> Useful Commands: #action #agreed #help #info #idea #link #topic.
18:19:00 <collabot> The meeting name has been set to 'cperf'
18:20:50 <dfarrell07> #info Hello World CBench job has been merged, help from Tim of Apex, not currently triggered but jamoluhrsen is working on it
18:23:39 <jamoluhrsen> #link https://git.opendaylight.org/gerrit/#/c/37639/ <--- patch with maybe relevant ovsdb scale test stuff
18:23:49 <dfarrell07> #info Nikos has been stuck doing internal stuff last two weeks, has been doing some prelem testing with multinet running in Docker containers, have been able to get mininet topos running with kernel-based OVS
18:24:03 <dfarrell07> #info Required tuning, like system file limits, to get mininet running
18:24:22 <dfarrell07> #info were able to get switch scale tests running with about 3k switches, 16 docker containers running multinet
18:24:39 <dfarrell07> #info up to 1.5k switches, perf seems to be same as VM test
18:25:07 <dfarrell07> #info may be problem with too few resources for containers, not an issue in kernel world, still digging
18:25:43 <dfarrell07> #info using VMs is an alternative of course, but doing container tests to figure out for sure if there's a scale issue in the kernel, to know if we really need VMs for kernel isolation
18:26:31 <dfarrell07> #info also working on NB traffic generation logic, to make traffic stateful, to track on which switches have flows installed, enable more complex scenarios like ones Al is talking about with topo failures
18:27:12 <dfarrell07> #info question about if ODL enables NB app to get notifications about if network changes, like if change in data store, LuisGomez says it should be possible using websockets
18:27:42 <dfarrell07> #info flows get updated through stat reports, so don't get exact time flows are installed or as up-to-date notifications
18:28:25 <dfarrell07> #info LuisGomez asks question to NIkos about when pushing tons of flows into OVS, it takes a bit to get stats/flow counts from OVS, how do they do that/are they seeing this delay
18:28:58 <dfarrell07> #info LuisGomez is trying to figure out if our test script is slow to collect this data or if OVS is just slow at reporting these stats
18:31:48 <dfarrell07> #info Detailed discussion about metrics collection and timing, tech details of implementation
18:32:52 <dfarrell07> #info Nikos doesn't think he's been collecting this number, LuisGomez asks him to check if possible
18:33:23 <dfarrell07> #info Programming 100k flows is super fast with bulk-o-matic, like 5s, but it also takes like 5s to get stats, so makes it difficult to measure metric
18:33:54 <dfarrell07> #info LuisGomez is interested in better ways to collect this data, if Nikos has such a way
18:34:21 <dfarrell07> #info LuisGomez asked this of Verex test people last week as well, no response so far
18:35:34 <dfarrell07> #info LuisGomez is looking at tests we want to run in next ODL whitepaper
18:35:51 <dfarrell07> #info One good one is REST NB flow programming
18:36:04 <dfarrell07> #info another is measuring flow programming on SB, LuisGomez needs to look into tooling for that
18:36:30 <dfarrell07> #info Nikos is measuring this type of thing with NSTAT already
18:37:23 <dfarrell07> #info dfarrell07 asks about Peter and Jan joining whitepaper efforts, Vratko says that Peter may be joining
18:38:48 <dfarrell07> #info Plan is to build tests on ODL infra, then copy robot logic to Intel lab to get numbers on HW
18:41:19 <dfarrell07> #info Cluster testing will not be a different test, there will be a test, and it will be done with single and cluster (if applies)
18:42:32 <dfarrell07> #info Current test are fairly black box, so should be about the same for 1 node or 3 or whatever
18:42:40 <dfarrell07> #info may run into differences if sending flows to leader or follower or whatever
18:43:08 <dfarrell07> #info dfarrell07 would like to bring cluster devs into this conversation, to get their feedback about tests
18:45:42 <dfarrell07> #info Discussion about data store vs RPC drop tests, we're not doing them in ODL because they are unstable, Nikos is doing DS test but is seeing lower numbers
18:14:38 <collabot> dfarrell07: Error: Can't start another meeting, one is in progress.  Use #endmeeting first.
18:14:42 <dfarrell07> #endmeeting