#opnfv-cperf: cperf
Meeting started by dfarrell07 at 18:14:46 UTC
(full logs).
Meeting summary
-
- http://testresults.opnfv.org/reporting/functest/release/master/index-status-apex.html
Functest status page for Apex, red because they don't have enough
runs I'm told (dfarrell07,
18:18:24)
- Al has an update on test strategy on CPERF-3
JIRA (dfarrell07,
18:22:19)
- https://build.opnfv.org/ci/view/cperf/job/cperf-apex-intel-pod2-daily-master/
cperf job running (jamoluhrsen,
18:22:20)
- https://jira.opnfv.org/browse/CPERF-3
JIRA to ID which metrics to target, making progress (dfarrell07,
18:22:26)
- dfarrell07 has been working adding NSTAT to
tools container, mostly done but not pushed, will push today or
weekend (dfarrell07,
18:22:34)
- Jamo has been making progress on getting
Jenkins job running on our pod with help from Tim, has the basics
running and is doing the normal debugging process now (x installed,
y configed, etc) (dfarrell07,
18:23:23)
- dfarrell07 had discussion on Test call about
layout of projects related to Functest vs Yardstick, where projects
should go. Functest for failures, Yardstick for things that gather
metrics. CPerf is under Yardstick. (dfarrell07,
18:24:17)
- OPNFV Summit is in two weeks, dfarrell07,
n_anastop, Al will be there (dfarrell07,
18:24:41)
- Marcus is still off doing OpenStack
training (dfarrell07,
18:24:48)
- Al discusses idea around having set of expected
things/flow we define as deployment (these switches, this # of
controllers, etc), param those numbers, work through normal usecases
vs driving to max capacity (dfarrell07,
18:25:55)
- LuisGomez and dfarrell07 agree with Al's
~user-story testing style idea, black box, maybe multinet test via
NSTAT (dfarrell07,
18:27:55)
- LuisGomez has been working on Bulk-o-Matic tool
for ODL. Runs inside controller but accessed via REST API, allows
user to say "program a ton of flows", then internal app drives
controller to do that from inside (dfarrell07,
18:28:47)
- LuisGomez reports that we get better perf with
Bulk-o-Matic test than network restart test we did in whitepaper (as
described last week, network restart has some unexpected
delay) (dfarrell07,
18:29:37)
- LuisGomez says better perf with Bulk-o-Matic
when programming more flows, but same when programming same number
of flows (200 vs 100k for example) (dfarrell07,
18:30:06)
- There is info on openflowplugin-dev list about
how to use this app, LuisGomez will share with ppl who don't follow
that list (dfarrell07,
18:30:27)
- Flow-config-blaster doesn't work well with more
than 200 flows from outside REST API (dfarrell07,
18:31:14)
- This Bulk-o-Matic tool maybe isn't good for
CPerf because it's not controller agnostic, would need one for each
controller (ONOS has a similar thing so that's maybe possible), but
again this kinda gets away from the point of common tests
(dfarrell07,
18:32:49)
- Agreement overall we should stick with more
controller agnostic tests (dfarrell07,
18:33:11)
- https://www.eventscribe.com/2016/OPNFV/aaSearchByDay.asp?h=Full%20Schedule&BCFO=P|G
OPNFV Summit Schedule (dfarrell07,
18:34:18)
- Al talks about about thinking of tests at
higher levels of abstraction, like a connectivity problem, and
letting controller (or ppl configuring it) solve that problem the
best way it can. Internal layer may be controller-specific, but
could still talk about black-box perf and compare (dfarrell07,
18:35:36)
- Could do something with switches already
created, then partition network and cause a lot of flows to be
programmed. Would be same task for all controllers, would be
agnostic via network. (dfarrell07,
18:38:04)
- LuisGomez talks about path programming test
that may also work for "authentic" test idea (dfarrell07,
18:38:29)
- n_anastop reports that his team has been
experimenting with Flow Config Blaster script (dfarrell07,
18:40:41)
- n_anastop there may be a perf flaw with the
test that's introducing extra overhead and may be causing numbers to
be reported worse than they actually are (dfarrell07,
18:41:51)
- n_anastop has sent a patch with a proposed
correction for this bug, we should review it (dfarrell07,
18:42:11)
- n_anastop has also been working on moving NSTAT
host to Docker containers, running each of the hosts in each
container, working with Docker compose to maybe solve this
(dfarrell07,
18:42:42)
- https://wiki.opendaylight.org/view/Openflow:Testing#Test_Case_3_-_Network_Failure_benchmark
(LuisGomez,
18:43:34)
- Al points out that we need to make sure network
between containers isn't a bottleneck, we should make sure of this
once built (dfarrell07,
18:45:38)
- n_anastop has also been working on latte, has
reduced overhead of packet decoding by 45 times, has been working on
special impl of decoding OF, go lib he's using doesn't have support,
needs access to various fields to impl test (dfarrell07,
18:47:17)
- Practice your German ;) (dfarrell07,
18:50:09)
Meeting ended at 18:50:12 UTC
(full logs).
Action items
- (none)
People present (lines said)
- dfarrell07 (32)
- collabot` (4)
- LuisGomez (1)
- jamoluhrsen (1)
Generated by MeetBot 0.1.4.