UPDATED Some networking companies see that they are losing control of the data center networks when it comes to blades and virtualization. One has reacted by making their own blades, others have come up with strategies and collaborating on standards to try to take back the network by moving the traffic back into the switching gear. Yet another has licensed their OS to have another company make blade switches on their behalf.
Where at least part of the industry wants to go is move the local switching out of the hypervisor and back into the Ethernet switches. Now this makes sense for the industry, because they are losing their grip on the network when it comes to virtualization. But this is going backwards in my opinion. Several years ago we had big chassis switches with centralized switch fabrics where(I believe, kind of going out on a limb here) if port 1 on blade 1 wanted to talk to port 2, then it had to go back to the centralized fabric before port 2 would see the traffic. That’s a lot of distance to travel. Fast forward a few years and now almost every vendor is advertising local switching. Which eliminates this trip. Makes things faster, and more scalable.
Another similar evolution in switching design was moving from backplane systems to midplane systems. I only learned about some of the specifics recently, prior to that I really had no idea what the difference was between a backplane and a midplane. But apparently the idea behind a midplane is to drive significantly higher throughput on the system by putting the switching fabric closer to the line cards. An inch here, an inch there could mean hundreds of gigabits of lost throughput or increased complexity/line noise etc in order to achieve those high throughput numbers. But again, the idea is moving the fabric closer to what needs it, in order to increase performance. You can see examples of a midplane systems in blades with the HP c7000 chassis, or in switches in the Extreme Black Diamond 20808(page 7). Both of them have things that plug into both the front and the back. I thought that was mainly due to space constraints on the front, but it turns out it seems more about minimizing the distance of connectivity between the fabric on the back and the thing using the fabric on the front. Also note that the fabric modules on the rear are horizontal while the blades on the front are vertical, I think this allows the modules to further reduce the physical distance between the fabric and the device at the other end by directly covering more slots, less distance to travel on the midplane.
Moving the switching out of the hypervisor, if VM #1 wants to talk to VM #2, having that go outside of the server and make a U-turn and come right back into it is stupid. Really stupid. It’s the industry grasping at straws trying to maintain control when they should be innovating. It goes against the two evolutions in switching designs I outlined above.
What I’ve been wanting to see myself is to integrate the switch into the server. Have a X GbE chip that has the switching fabric built into it. Most modern network operating systems are pretty modular and portable(a lot of them seem to be based on Linux or BSD). I say integrate it onto the blade for best performance, maybe use the distributed switch frame work(or come up with some other more platform independent way to improve management). The situation will only get worse in coming years, with VM servers potentially having hundreds of cores and TBs of memory at their disposal, your to the point now practically where you can fit an entire rack of traditional servers onto one hypervisor.
I know that for example Extreme uses Broadcom in most all of their systems, and Broadcom is what most server manufacturers use as their network adapters, even HP’s Flex10 seems to be based on Broadcom? How hard can it be for Broadcom to make such a chip(set) so that companies like Extreme (or whomever else might use Broadcom in their switches) could program it with their own stuff to make it a mini switch?
From the Broadcom press release above (2008):
To date, Broadcom is the only silicon vendor with all of the networking components (controller, switch and physical layer devices) necessary to build a complete end-to-end 10GbE data center. This complete portfolio of 10GbE network infrastructure solutions enables OEM partners to enhance their next generation servers and data centers.
Maybe what I want makes too much sense and that’s why it’s not happening, or maybe I’m just crazy.
UPDATE – I just wanted to clarify my position here, what I’m looking for is essentially to offload the layer 2 switching functionality from the hypervisor to a chip on the server itself. Whether it’s a special 10GbE adapter that has switching fabric or a dedicated add-on card which only has the switching fabric. Not interested in offloading layer 3 stuff, that can be handled upstream. Also interested in integrating things like ACLs, sFlow, QoS, rate limiting and perhaps port mirroring.