Darren Reed
2008-Apr-18 08:37 UTC
[crossbow-discuss] Does the virtual networking need to be ethernet?
Nicolas (and others), On thinking about the vnic devices and virtual switches that we''re looking to put into Solaris with crossbow, I wondered, do we need to emulate ethernet internally? So let me expand on that a little... For VNICs that sit atop physical NICs, there is no choice - it has to have the same MAC as the NIC below it. But for VNICs that connect internally and only to the vswitch, what use is an ethernet header if there is never any exchange between a real ethernet? And if a non-ethernet MAC header is in place, what sort of performance benefits can be had from not having to worry about header alignment (for example)? The challenge to this idea would be does it make sense to design and introduce something new that still has to provide multicast and broadcast behaviour (at least so that ARP works.) Darren
Nicolas Droux
2008-Apr-18 20:46 UTC
[crossbow-discuss] Does the virtual networking need to be ethernet?
On Apr 18, 2008, at 2:37 AM, Darren Reed wrote:> Nicolas (and others), > > On thinking about the vnic devices and virtual switches > that we''re looking to put into Solaris with crossbow, I > wondered, do we need to emulate ethernet internally? > > So let me expand on that a little... > > For VNICs that sit atop physical NICs, there is no > choice - it has to have the same MAC as the NIC > below it. > > But for VNICs that connect internally and only to > the vswitch, what use is an ethernet header if there > is never any exchange between a real ethernet?An advantage of using ethernet is that it enables us to "easily" and efficiently bridge a virtual network with a physical network.> And if a non-ethernet MAC header is in place, > what sort of performance benefits can be had from > not having to worry about header alignment (for > example)?There are local optimizations that we have on our list to optimize data-paths between MAC clients sharing the same underlying NIC or etherstub. These optimizations include straight-through LSO/LRO, and outbound classification bypass for local destinations. Once we have these optimizations in place, the cost of having to deal with MAC headers themselves should be much reduced, and hopefully negligible :-)> The challenge to this idea would be does it make > sense to design and introduce something new that > still has to provide multicast and broadcast > behaviour (at least so that ARP works.)Right. I think it would be desirable to use ethernet for the reasons I listed above, and because users/admins/tools already know how to deal with ethernet. Nicolas.> > > Darren > > _______________________________________________ > crossbow-discuss mailing list > crossbow-discuss at opensolaris.org > http://mail.opensolaris.org/mailman/listinfo/crossbow-discuss-- Nicolas Droux - Solaris Networking - Sun Microsystems, Inc. nicolas.droux at sun.com - http://blogs.sun.com/droux