#opendaylight-group-policy: Renderer
Meeting started by alagalah at 16:09:09 UTC
(full logs).
Meeting summary
- renderer_common (regXboi, 16:14:28)
- https://wiki.opendaylight.org/view/Group_Policy:Architecture/Renderer_Common
(regXboi,
16:14:45)
- lenrow asks what is tenant (regXboi,
16:15:00)
- readams answers that it is the top level
container for policies (regXboi,
16:15:22)
- plan is to build two renderers: OpenFlow and
OpFlex (tbachman,
16:15:30)
- WIll use OVS-based overlay (tbachman,
16:15:40)
- lenrow points out that this is an overloaded
term across ODL and so it might change (regXboi,
16:15:45)
- https://wiki.opendaylight.org/view/Group_Policy:Architecture/OVS_Overlay
Description of overlay network (tbachman,
16:16:55)
- Overlay needs MTU that can support 1500 byte
packets (e.g. 1600 MTU, to account for tunnel encap overhead)
(tbachman,
16:17:48)
- regXboi notes that this is a real issue
(tbachman,
16:17:55)
- The overlay model is simple, in order to be
able to demonstrate use of GBP (tbachman,
16:19:11)
- lenrow asks if this design has an OpFlex
connection to one OVS instance and OpenFlow connection to the other
OVS instanct (tbachman,
16:20:18)
- This currently isn’t planned — will be either
or, but not both (tbachman,
16:20:34)
- The Overlay example has two subnets
(tbachman,
16:21:06)
- it’s important to have a logically “out of
band" control channel (tbachman,
16:21:50)
- The example uses VxLAN encap (tbachman,
16:22:04)
- VxLAN encap has network identifier (tbachman,
16:22:49)
- would allocate a universal gateway map, which
would be owned by distributed router (tbachman,
16:23:25)
- simple lookup to see if L2 is targeting
GW (tbachman,
16:23:37)
- For bridging, deliver to destination
(tbachman,
16:23:52)
- For routing, if destination is on a remote
vSwitch, it’s 2 hops: one on-local vSwitch, another on the
destination vSwitch, which performs correct IP address to MAC
mapping (tbachman,
16:24:39)
- Determination of Next Hop destination for any
given EP is renderer-specific (tbachman,
16:25:22)
- for OpFlex, will likely be reactive, since
agent is on the switch (tbachman,
16:25:35)
- for OpenFlow, will likely be a hybrid
(tbachman,
16:25:46)
- Question on reactive in OpFlex: driven by the
controller or agent? (tbachman,
16:26:09)
- For learning, this will be a packet-in
(tbachman,
16:26:19)
- controller then pushes this information where
needed (tbachman,
16:26:48)
- dvorkinista notes that this is a
proof-of-concept (tbachman,
16:28:06)
- regXboi notes that this is rebuilding
OpenDove (tbachman,
16:29:20)
- dvorkinista asks if we can leverage
OpenDove (tbachman,
16:29:37)
- Since this is a proof-of-concept for GBP, this
might make sense (tbachman,
16:30:14)
- regXboi says there are some slight differences,
but with so it might be good to leverage pieces (tbachman,
16:30:42)
- lenrow notes that OpenDove is currently
competing with GBP, so bringing rending code into GBP would be
great (tbachman,
16:31:11)
- dave_tucker notes that OVSDB also has
netvirt (tbachman,
16:31:49)
- lenrow adds VTN to that (tbachman,
16:31:55)
- regXboi says that we probably should look at
OVSDB, since it’s the most mature (tbachman,
16:32:14)
- AGREED: opendaylight-ovsdb guys are “mature” ;) (tbachman,
16:32:36)
- lenrow points out that it’s important to get
all things writing to flow tables under the GBP umbrella
(tbachman,
16:33:28)
- regXboi has an issue with that, but that should
be taken up separately (tbachman,
16:33:45)
- dvorkinista notes that GBP does not dictate who
gets to write renderers. (tbachman,
16:34:06)
- where projects can decide if they want to have
plugins that participate as part of GBP (tbachman,
16:34:26)
- Question for OpFlex renderer case: there is an
agent needed for the OVS switch (tbachman,
16:35:25)
- dvorkinista says yes, and the interface between
the agent and OVS is OpenFlow and OVSDB (tbachman,
16:35:45)
- Since routing and bridging is happening, that
means we’re injecting L2 and L3 adjacencies when a new host comes
up (tbachman,
16:36:50)
- Question of whether it’s worthwhile to support
L2, and not just do L3 (tbachman,
16:37:12)
- readams notes that you can do that, but that
some folks like L2 as well (tbachman,
16:37:36)
- One tricky part is allowing hosts to
communicate with the outside world (tbachman,
16:37:58)
- There are two broad approaches, and not clear
which one this proof of concept will use (tbachman,
16:38:25)
- First approach is OVS gateway boxes and map an
external endpoint onto one of the external interfaces on that box,
and then use that box as a routing bridge between the overlay
network and the outside world (tbachman,
16:38:54)
- That has the advantage in that it’s
conceptually simple (i.e. easy to conceive) (tbachman,
16:39:13)
- The other approach is a dNAT scheme
(tbachman,
16:39:25)
- For any given switch, there’s a pool of IPs
that it can allocate, and any endpoint that wants to communicate
wtih the outside world gets an IP from that pool. (tbachman,
16:40:17)
- current plan is to use dNAT scheme (tbachman,
16:40:47)
- lenrow asks who’s going to be impacted by
configuration complexity? (tbachman,
16:41:11)
- readams says it’s an operational impact
(tbachman,
16:41:20)
- regXboi says that dNAT doesn’t scale, b/c
scaling pools with each vSwitch (tbachman,
16:43:09)
- dvorkinista emphasizes this is proof of
concept, so this is probably fine (tbachman,
16:43:49)
- regXboi notes that OpenDove does have a GW, but
that OpenDove is not participating in Helium, b/c it wants pieces in
place (e.g. GBP) before adding that feature. (tbachman,
16:47:25)
- the OVSDB project may have pieces that can be
leveraged for the renderer (tbachman,
16:53:59)
- The actual OF renderer itself isn’t yet flushed
out. (tbachman,
16:54:16)
- https://wiki.opendaylight.org/view/Group_Policy:Architecture/OpenFlow_Renderer
Description of what OF renderer needs (tbachman,
16:54:57)
- lenrow asks if enhancing some components like
topology manager for lookup can be leveraged for this example
(tbachman,
16:58:15)
- readams says that we of course could/should
leverage existing controller components (tbachman,
16:58:33)
- mickey asks whether we are merging mapping
concepts with the endpoint registry (tbachman,
17:00:18)
- readams notes that dvorkin says this should be
handled in the EP registry, using context (tbachman,
17:01:04)
- Where a field exists that isn’t defined by the
model, but renderers know how to interpret and use (tbachman,
17:01:35)
- regXboi is over on model hangout keeping dconde
company.... (regXboi,
17:03:18)
- alagalah: you need to herd folks to the new
hangout (regXboi,
17:03:44)
- I am here in the hangout with regXboi
(dconde,
17:03:56)
- Renderer meeting bleeding into model
meeting (tbachman,
17:05:39)
Meeting ended at 17:10:43 UTC
(full logs).
Action items
- (none)
People present (lines said)
- tbachman (72)
- regXboi (18)
- odl_meetbot (7)
- alagalah (7)
- dconde (2)
- mestery (1)
- dave_tucker (1)
Generated by MeetBot 0.1.4.