Open Stack

Hi Mathieu,

Thanks for the email. As discussed during the ML2 IRC meeting [2], we have
not decided on a design. That is why we do not have a spec for review yet.
The idea is that we spend a bit more time and figure out the details and
try out some possible options before we go ahead with the spec. So new
comments/suggestions are much appreciated.

In addition to having different drivers we want to reduce the code
replication across current agents. I am wondering if with what you are
proposing as dataplane drivers, we will end up with having different
drivers which look like the current agents and we do not deal with reducing
code replication across agents. If this is not a correct assessment could
you describe how we can avoid code replication across agents/drivers.

Let me briefly explain what I have outlined in [3] (also mentioned in [2]).
We are thinking of having drivers for each extension or probably better
said each functionality. So we can have a base l2 connectivity driver, an
l2pop driver, a sg driver (not to be confused with sq drivers), so on so
forth. I think in your email you are referring to these drivers (or
something close to them) as Extension drivers. In [3] they are called Agent
Drivers.

Then we have the Resource Drivers which will be essentially used for
realizing these features depending on the technology/resource being used
(e.g., using  OVS switches, or Linux Bridges, or some other technology).
The main reason for using such a organization is to be able to have
different agent drivers utilize the same resource and reuse code. The
challenge is figuring out the api for such a driver. Any thoughts on this?

Mohammad

[3] https://etherpad.openstack.org/p/modular-l2-agent-outline




From:	Mathieu Rohon <mathieu.rohon at gmail.com>
To:	OpenStack Development Mailing List
            <openstack-dev at lists.openstack.org>, Mohammad
            Banikazemi/Watson/IBM at IBMUS,
Date:	05/30/2014 06:25 AM
Subject:	[openstack-dev][Neutron][ML2] Modular agent architecture



Hi all,

Modular agent seems to have to choose between two type of architecture [1].

As I understood during the last ML2 meeting [2], Extension driver
seems to be the most reasonnable choice.
But I think that those two approaches are complementory : Extension
drivers will deal with RPC callbacks form the plugin, wheras Agent
drivers will deal with controlling the underlying technology to
interpret those callbacks.

It looks like a controlPlane/Dataplane architecture. Could we have a
control plane manager on which each Extension driver should register
(and register callbacks it is listening at), and a data plane manager,
on which each dataplane controller will register (ofagent, ovs, LB..),
and which implement a common abastract class.
A port will be managed by only one dataplane controller, and when a
control plane driver wants to apply a modification on a port, it will
retrieve the correct dataplane controller for this port in order to
call one of the abstracted method to modify the dataplane.


[1]
https://wiki.openstack.org/wiki/Neutron/ModularL2Agent#Possible_Directions
[2]
http://eavesdrop.openstack.org/meetings/networking_ml2/2014/networking_ml2.2014-05-28-16.02.log.html


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140530/c737a1d8/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: graycol.gif
Type: image/gif
Size: 105 bytes
Desc: not available
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140530/c737a1d8/attachment.gif>