#opendaylight-group-policy: model/arch redux
Meeting started by regXboi at 18:16:36 UTC
(full logs).
Meeting summary
-
- Is creating a tenant in response to a new flow
a good idea? answer: no (tbachman,
18:17:18)
- however, we still need a bulk operation to
create tenants (tbachman,
18:17:33)
- orchestration systems (openstack) allows for
the bulk creation of networks (tbachman,
18:18:23)
- readams asserts that's fine (tbachman,
18:18:28)
- tenant/EPG changes shouldn't be continuous
operations (tbachman,
18:19:20)
- regXboi notes the M&A use case, which
implies a running system, and a new bulk has been dumped on
it (tbachman,
18:20:06)
- time constant for convergence should be
relatively short (tbachman,
18:22:29)
- regXboi would like to see the time constant for
convergence on the order of seconds (tbachman,
18:24:37)
- and by time constant, that's the "1/e
time" (tbachman,
18:24:56)
- There are situations where a renderer running
on physical switches running OF and creating L2 flows of 10k/second,
and 50 switches, it's possible that there could be changes in policy
that could change flows everywhere, making for a convergence time on
the order of minutes (tbachman,
18:26:24)
- but those are seen as "rare" occurrences
(tbachman,
18:26:38)
- It would be good if we could make some
statement on the convrgence time of some cases. (tbachman,
18:27:21)
- regXboi asks what call rate they're looking at
for UC&C use case (tbachman,
18:31:42)
- uchau says about 200 flows/second, 10-100k
sessions active at any one time (tbachman,
18:31:59)
- correction 200 sessions/second (tbachman,
18:32:17)
- alagalah asks for clarification of 10-100k
number -- active, or total possible number (tbachman,
18:33:51)
- Actually what I said was clarify 10k-100k
concurrent sessions or that number of endpoints (alagalah,
18:35:03)
- lost audio (tbachman,
18:35:05)
- clarification on what a dynamic classifier
is (tbachman,
18:36:04)
- The config says this is the description of how
these endpoints in different EPGs are allowed to communicate
(tbachman,
18:39:02)
- It's not a rate of change number that separates
operational vs. config changes. (tbachman,
18:42:52)
- subject feature definition is a global
configuration item (tbachman,
18:43:37)
- A rule is defined that consists of a classifier
that's associated with the flow (tbachman,
18:44:29)
- the action would be "allow", "apply QoS"
(tbachman,
18:44:40)
- The renderer gets a packet_in, it goes through
policy resolution, discovers that it needs to classify with the lync
flow classifier (tbachman,
18:46:12)
- readams recommends reading the policy
resolution spec on the wiki (tbachman,
18:46:28)
- resolve the EPG of the EP, using the EPG
repo (tbachman,
18:46:36)
- https://wiki.opendaylight.org/view/Group_Policy:Architecture/Policy_Model
(tbachman,
18:47:21)
- https://wiki.opendaylight.org/view/Group_Policy:Architecture
(tbachman,
18:47:33)
- there is something like an orchestration system
that assigns EPs to EPGs (tbachman,
18:48:34)
- regXboi makes Lync call to self (tbachman,
18:50:02)
- policy was improperly specified (tbachman,
18:50:14)
- renderer gets information, ensures that EP gets
assigned to proper EPG (tbachman,
18:50:41)
- There's some mechanism for assigning/mapping an
EP to it's EPG (tbachman,
18:51:50)
- regXboi has 10 more minutes before his next
meeting (regXboi,
18:51:51)
- Lync flow classifier has something backing it
-- the Lync session info (tbachman,
18:52:47)
- The point is that the operational data can be
treated differently than the configuraiton data. (tbachman,
18:54:25)
- there's going to be auditing, AAA, logging,
etc. (tbachman,
18:55:15)
- Still a question of how an EP is mapped to an
EPG -- readams says this is a longer conversation, but there's got
to be some configuration of your network that does that (tbachman,
19:01:00)
- For example, in UC&C case, the devices
making the calls can be the EPs (tbachman,
19:03:59)
- Lync session server pushes information about
lync session into some storage, which triggers an event to the
renderer (tbachman,
19:05:39)
- what's in the policy is how does the
application want to treat these flows (tbachman,
19:06:02)
- FOr example, only allow calls to other parties
in a specific region (tbachman,
19:06:25)
- the renderer listens to the policy repository
and to the store that keeps the sessions (tbachman,
19:06:46)
- uchau expresses concern about having a
per-application renderers (tbachman,
19:08:35)
- readams notes that the renderers could be
generalized -- e.g. campus LAN renderer (tbachman,
19:08:49)
- the goal for UC&C is to support campus,
enterprise, and datacenter (tbachman,
19:16:15)
- the ONF is wondering if GBP can support the use
cases they're looking at (tbachman,
19:17:03)
- ACTION: alagalah to
send a diagram of what this may look like (tbachman,
19:20:35)
- ACTION: uchau to
provide link for UC&C (tbachman,
19:20:42)
- the policy model is extremely general, but the
first implementation is OpenStack focused (tbachman,
19:21:36)
Meeting ended at 19:25:41 UTC
(full logs).
Action items
- alagalah to send a diagram of what this may look like
- uchau to provide link for UC&C
Action items, by person
- alagalah
- alagalah to send a diagram of what this may look like
- uchau
- uchau to provide link for UC&C
People present (lines said)
- tbachman (57)
- regXboi (13)
- odl_meetbot (4)
- alagalah (3)
- uchau (3)
Generated by MeetBot 0.1.4.