[ovs-dev] [OVS-dev]: OVN: RFC re: logical and physical endpoint separation proposal

Mickey Spiegel emspiege at us.ibm.com
Wed Mar 2 11:00:08 UTC 2016


I was wondering if Justin's proposal below would be supported. I did not see the proposed physical.c changes in the patches that Darrell sent out, so I was not sure if this would be the case or not. Justin's proposal does seem to preserve current behavior while adding some new behavior.

I do have one major concern regarding applicability of a physical endpoint binding to the software gateway proposal. I thought the software gateway proposal would be primarily aimed at L3, handling features like Floating IP and SNAT. When there are a large number of tenants that need the same external access (e.g. internet connectivity from a public data center), even when there is only one router per tenant with its gateway interface pinned to a chassis, the total number of router gateway interfaces sharing the same external network may be large. We would want the ability to spread different tenant router gateway interfaces across different chassis that are all connected to the same external network. In this case, one "localnet" port would still map to N chassis.

Mickey

-----Justin Pettit <jpettit at ovn.org> wrote: -----
To: Russell Bryant <russell at ovn.org>
From: Justin Pettit <jpettit at ovn.org>
Date: 03/01/2016 08:30PM
Cc: Mickey Spiegel/San Jose/IBM at IBMUS, "dev at openvswitch.org" <dev at openvswitch.org>, Darrell Lu <dlu998 at gmail.com>
Subject: Re: [ovs-dev] [OVS-dev]: OVN: RFC re: logical and physical endpoint separation proposal

> On Mar 1, 2016, at 6:44 PM, Russell Bryant <russell at ovn.org> wrote:
> 
> FWIW, this is what I was trying to figure out with my questions as well.
> It does seem like there is something missing here.
> 
> With localnet ports today, a single logical port maps to a physical port on
> N chassis.  It's not 1 to 1, which this model seems to assume.

Is that true?  I think the documentation implies that, but I'm not sure if the implementation does.  In the Logical_Port table, there can be zero or one "phys_endpt".  If someone doesn't populate "phys_endpt" (or sets "type" to "vlan" and "ingress_encap" to the vlan id), then it should have the behavior that you describe.  I spoke with Darrell at some length today, and I think he'll be sending out a revised version.

I think the final version is supposed to behave like this:  If "phys_endpt" is empty or exists with only a vlan tag, then it will behave just like it does now: the bindings will happen on any hypervisor with the appropriate "external-ids:ovn-bridge-mappings" set.  If "phys_endpt" specifies a particular chassis, then it would only be instantiated there.  (I think this latter case could be useful when creating a software gateway.)

I need to read the localnet implementation and Darrell's proposed changes more carefully, but does the previous paragraph sound reasonable?  I'll dig into it more tomorrow, but please let me know if I'm way off course.

--Justin







More information about the dev mailing list