#opendaylight-meeting: tws
Meeting started by colindixon at 17:00:10 UTC
(full logs).
Meeting summary
- agenda bashing (colindixon, 17:00:59)
- https://wiki.opendaylight.org/view/Tech_Work_Stream:Main#Upcoming_Meeting_Agendas
(colindixon,
17:01:29)
- if people have ideas for future meetings, feel
free to post them there or bring them up during the beginning/end of
TWS calls (colindixon,
17:01:57)
- the other topic for today is maros and jmedved
on the performance of NETCONF (colindixon,
17:05:37)
- netconf performance (colindixon, 17:09:58)
- netconf scale testing (colindixon,
17:11:23)
- connecting to 10,000 devices, it took about 15
minutes to connect to every device, then 5-6 inutes to make an RPC
call to each device (colindixon,
17:12:13)
- all tests using 4 core CPU with 6G of RAM for
ODL (colindixon,
17:12:23)
- the above was with the clustered data
store (colindixon,
17:12:36)
- with the in-memory data store, it took ~13
minutes (vs. 15 to do the same thing) (colindixon,
17:12:55)
- can mount 12-12.5k devices total with 4G of
RAM (colindixon,
17:13:31)
- colindixon asks where these scripts are
running, maros says so far private, but working on getting them into
integration (colindixon,
17:13:57)
- netconf peformance testing (colindixon,
17:14:09)
- netconf-test-client => netconf server in ODL
=> MD-SAL’s data store (in-memory) (colindixon,
17:14:26)
- running on 16 core, 32G RAM machine
(colindixon,
17:14:34)
- data writes are simple L2-FIB-style data
(colindixon,
17:14:58)
- without baching, we get to 2400 writes/sec sync
and 5600 writes/sec async (colindixon,
17:15:31)
- could get more than 40,000 writes/sec with
async and batching 1000 writes/sec (colindixon,
17:16:04)
- jmedved points out that this is vs. ~1000-1500
writes/sec in OpenFlow (colindixon,
17:17:02)
- colindixon asks why that’s the case since they
should both be just writing to the data sore (colindixon,
17:17:19)
- jmedved says part of that was that with
openflow it’s a write, read, and then write, with batching the
number goes 10-12k writes per second even with openflow, which
begins to look not too dissimilar (colindixon,
17:18:15)
- with multiple clients and batching, it goes as
high as 130,000+ writes/second (colindixon,
17:18:36)
- how many L2-FIB-link entries could be stored in
limited memory something like 1.4 million decreasing to 300k as you
go from 2G RAM to 500M RAM (colindixon,
17:19:34)
- netconf notification performance (colindixon,
17:19:46)
- notifications are basically a summary of a few
routes (colindixon,
17:20:11)
- notifications/second range from 3k
notifications/second to 22k notifications/second (increase due to
smaller size and reduced complexity) (colindixon,
17:21:17)
- end-to-end netconf performance (colindixon,
17:21:28)
- restconf-test-client => restconf =>
MD-SAL binding aware app => netconf connector => netconf
device (colindixon,
17:21:59)
- simulated netconf device was parsing data and
throwing it way (colindixon,
17:22:26)
- data was also a list of routes (colindixon,
17:22:33)
- ranging from 600 routes/second without
batching, multiple clients, or asynchrony to 6000 routes/second with
things tuned up (colindixon,
17:23:19)
- this went up to 180,000 routes/second when
moving to something avoiding restconf, restconf seems to be what’s
slowling things down (colindixon,
17:24:00)
- jmedved asks when memory constrained, what was
the peformance, maros says it was pretty fast until it got to ~90%
of capacity (colindixon,
17:24:24)
- jmedved points out that this would be useful
for the honeycomb ideas (colindixon,
17:24:37)
- colindixon asks about memory efficiency, it
looks like we use ~1.5G RAM to store ~1,000,000 extra L2 FIB
entries, which is ~1.5KB/entry, which isn’t great, but probably OK
of Java (colindixon,
17:27:26)
- jmedved asks rovarga what could be done to
further optimize memory usage, rovarga said we could use custom data
structures or wait for Java 9 and some of it’s features (colindixon,
17:28:15)
- points out we should look into this more deeply
for honeycomb and what the trade-offs (colindixon,
17:31:07)
- there are some discussions about the relative
reasons for openflow vs. netfconf performance differences
(colindixon,
17:32:06)
- LuisGomez asks if it will get into ODL CI,
Maros says yes that’s the plan (colindixon,
17:32:25)
- Art says he’d like to put out tutorials of how
NETCONF would work so he can help out, would like somebody to reach
out to him (colindixon,
17:34:17)
- jmedved and Art will connect (colindixon,
17:34:30)
- user stories (colindixon, 17:34:39)
- https://wiki.opendaylight.org/view/UserStories:Main
(julim,
17:35:16)
- https://wiki.opendaylight.org/view/UserStories:Main
the user stories we’re going to go through (colindixon,
17:35:29)
- julim says that this is a list of user stories
around OpenDaylight and OpenStack primarily, there are a variety of
roles defined (colindixon,
17:36:58)
- each user story is a list of scenarios that are
either not yet covered by opendaylight or only partially
covered (colindixon,
17:37:41)
- nyechiel says these are the basic table stakes
for working with OpenStack (colindixon,
17:37:58)
- https://wiki.opendaylight.org/view/File:ODL_User_Stories.pdf
this is the full backlog of user stories that they have put together
(colindixon,
17:38:14)
- edwarnicke asks about the first one, he was
under the impression that security groups were working (colindixon,
17:39:32)
- edwarnicke asks they to check with group-based
policy working (colindixon,
17:39:49)
- nyechiel points out that so far this is mostly
looking at OVSDB, no GBP, but he’ll look at that (colindixon,
17:40:10)
- edwarnicke notes that he’s pretty sure that GBP
supports security groups and IPv6 security groups fully
working (colindixon,
17:40:42)
- edwarnicke believes everything here is working
insofar as he knows except for DHCP/IPAM and conntracking features
that require new OVS (colindixon,
17:43:15)
- nyechiel and julim say that this isn’t so much
about these items, although they might be useful, but more that this
is likely a better model for how to set priorities and focus in
OpenDaylight (colindixon,
17:43:55)
- nyechiel and julim also point out that if we
want tration with the openstack community (and their users) this is
likely a good approach (colindixon,
17:44:40)
- https://wiki.opendaylight.org/view/UserStories:Explanation
put some really basic information about how to appraoch things with
user stories (courtesy of julim) (colindixon,
17:45:24)
- Prem_ asks how this changes with the big tent
approach in OpenStack (colindixon,
17:45:45)
- nyechiel says that’s a longer discussion, but
he doesn’t thing we’re far enough along in our user story planning
for that to dramatically affect us (colindixon,
17:46:24)
- art says he really likes this, but he’d love to
see this drift even further up the stack from feature-level things
to even more use cases (colindixon,
17:50:05)
- julim says that soudns great, and ask for any
next steps (colindixon,
17:50:16)
- art says he is trying to come up with a first
stab at well-define use caes and is hoping for feedback soon
(colindixon,
17:50:35)
- dneary asks if there are any high-priority use
cases missing from here? (colindixon,
17:51:05)
- edwarnicke asks about NFV use cases, he sees
NFV as the end-game and OpenStack is just stepping stone
(colindixon,
17:52:22)
- edwarnicke says one idea about how to be more
story/user-oriented would be to make sure we at least acknowledge
when we do meet them (colindixon,
17:55:42)
- vishnoianil__ says that we have a broad set of
good ideas and ways forward for things, but we often don’t end up
with effective places to discuss the ideas and requirements in a
single place (colindixon,
17:56:28)
- colindixon assuems that added onto
vishnoianil__’s comments is the idea that we could attach user
stories to it (colindixon,
17:56:55)
- edwarnicke says that odlforge would help a lot
as having a way to place code somewhere that could be played
with (colindixon,
17:57:18)
- phrobb says that these other comments point to
a broader thing about how to break down what we’re doing from a
user-oriented, requirements-driven, approach for features
(colindixon,
18:00:30)
- art asks how we do feature requests today,
colindixon says maybe that would be a good topic for a future
call (colindixon,
18:02:00)
Meeting ended at 18:02:22 UTC
(full logs).
Action items
- (none)
People present (lines said)
- colindixon (87)
- julim (19)
- odl_meetbot (5)
- nyechiel (4)
- edwarnicke (2)
- rovarga (1)
- phrobb (0)
- tbachman (0)
Generated by MeetBot 0.1.4.