
OpenFlow's power is that it lets an external controller tell switches and routers what to do. Those instructions are stored in ternary CAMs (TCAMs), a type of memory that switches and routers just happen to have.
When you apply OpenFlow to an ordinary switch, the TCAMs do the job, but not very impressively, Proch says. The recently announced G8264 from IBM includes 48 10Gbit/s ports -- hardly a commodity box -- but it supports less than 100,000 OpenFlow flows. "We've done research that says a 10Gbit/s Ethernet port might have 2.5 million flows," Proch says.
Netronome's processors can support 16 million flows. So what you'd do, he's suggesting, is send traffic to different chips depending on need. Traffic that needs to be treated as a flow would go to the NFP. Simpler traffic would be handled by an x86 processor or even the Ethernet switch fabric itself.
That sounds like a lot of chips, but Netronome uses that same arrangement for the normal line cards it designs. Cards were the company's original business, before buying the rights to flow-processor technology from Intel, so in a sense its OpenFlow pitch is just business as usual. (See Netronome Expands Its Mission.)
Proch isn't saying every OpenFlow switch has to be built this way, but if OpenFlow is going to support very large data centers, these kinds of designs would be relevant.
The commodity idea, by the way, comes from the fact that OpenFlow is designed to not care about hardware. For a lot of people, including me, that brings up a first impression of cheap, tissue-thin Ethernet switches that get thrown away like Twinkie wrappers. But that's not how it's going to be. Ethernet switches can still be interesting in an OpenFlow world.
I don't think OpenFlow will completely topple Cisco Systems Inc. (Nasdaq: CSCO), either, although that's a separate debate, one Cisco will grapple with as it tries to embrace software defined networking.
I bring up Cisco because I'm not going to see Proch's presentation Wednesday. It conflicts with Cisco's earnings announcement. There's a joke in there somewhere.
— Craig Matsumoto, Managing Editor, Light Reading
Both share the same architecture vision of centralized control plane.
QFabric is in backing for multi-years now. I wondering what is it taking them to roll-out completely. Scaling issues?
May be they should switch to OpenFlow based architecture. Oh, wait...OpenFlow is still in hype cycle, never mind.