[ovs-discuss] OVS with Distributed Data Plane

[Toth, Andrew]

That's pretty much how we're approaching it, though we're looking at a DPDK implementation so we're adapting the dpif-netdev implementation instead of the dpif-netlink. Thanks for the feedback.

-----Original Message-----
From: Ben Pfaff [mailto:blp at ovn.org] 
Sent: Monday, January 04, 2016 1:58 PM
To: Toth, Andrew <Andrew.Toth at stratus.com>
Cc: 'Justin Pettit' <jpettit at ovn.org>; discuss at openvswitch.org
Subject: Re: [ovs-discuss] OVS with Distributed Data Plane

I don't know anyone doing that, but it would only require writing new dpif and netdev implementations that talked across an RPC channel to a remote datapath.  You could probably adapt dpif-netlink to handle it by just using Netlink "wire format" across, say, TCP instead of Netlink.

On Mon, Jan 04, 2016 at 06:43:19PM +0000, Toth, Andrew wrote:
> We are looking to spawn the fast and slow paths on different VMs on separate hypervisors, not simply on different processors on the same hypervisor. I have not seen any mention of this type of set up around so I've started decomposing OVS in order to enable this functionality. Please update me if this functionality does exist somewhere, I'd love to save myself the heaps of work it appears to entail.
> 
> -----Original Message-----
> From: Ben Pfaff [mailto:blp at ovn.org]
> Sent: Monday, January 04, 2016 1:25 PM
> To: Toth, Andrew <Andrew.Toth at stratus.com>
> Cc: 'Justin Pettit' <jpettit at ovn.org>; discuss at openvswitch.org
> Subject: Re: [ovs-discuss] OVS with Distributed Data Plane
> 
> It's pretty common to run the OVS fast path and slow path on different processors.  You just set the affinity of each thread appropriately.
> 
> On Mon, Dec 28, 2015 at 12:37:40PM +0000, Toth, Andrew wrote:
> > Not as far as I can tell. My understanding is that OVS is built to support multiple data path types but not have the CE on one system and the FE on another. 
> > 
> > -----Original Message-----
> > From: Ben Pfaff [mailto:blp at ovn.org]
> > Sent: Wednesday, December 23, 2015 4:27 PM
> > To: Toth, Andrew <Andrew.Toth at stratus.com>
> > Cc: 'Justin Pettit' <jpettit at ovn.org>; discuss at openvswitch.org
> > Subject: Re: [ovs-discuss] OVS with Distributed Data Plane
> > 
> > On Tue, Dec 22, 2015 at 09:01:10PM +0000, Toth, Andrew wrote:
> > > Regrettably OVN will not fit our situation. In the NFV world 
> > > today, there is a push to separate the fast path data plane 
> > > processing from the slower path configuration control processing. 
> > > The common understanding is to put the data path processing on its 
> > > own processor as a separate process (possibly multiple data plane 
> > > processes on multiple processors) with cached data (flows in the 
> > > emc in the case of
> > > OVS) and communicate back to the control plane data store when the 
> > > required information is not in the cache. For our current project 
> > > we need an open source VNF that is decomposed in this way. I just 
> > > want to ensure that this separation is not already being done by 
> > > the OVS community before we go through the task of separating OVS 
> > > into these two components for our scenario.
> > 
> > Isn't OVS already built this way?