Juniper Networks was one of the first to challenge Cisco with an Ethernet switching fabric, announcing the ambitious QFabric architecture in February. At this week's Interop 2011 in Las Vegas, a UBM TechWeb event, it was joined by many other rivals launching or demonstrating Ethernet fabrics of their own, while the departure of QFabric architect David Yen for Cisco left some questions as to its future strategy.
Like other fabrics, the basic idea behind Juniper's is to make the network flatter, cutting out the rigid tree architecture in favor of more meshed links, offering the flexibility needed to support virtualization and enterprise clouds. The vision is to treat the network (or at least the fabric core) as one giant switch, so most vendors are focusing on aggregating switches together.
Juniper is no exception, but it says that its network really is a switch. According to Abner Germanow, Juniper's director of enterprise marketing, QFabric takes the components of a data center switch and massively expands them into separate components. "What used to be a line card looks like a top-of-rack switch, what used to be the backplane now looks like a chassis, and what used to be the supervisor is now a server," he said in an interview."
According to Juniper, replacing the interface between "line card and backplanes with a network link doesn't create a bandwidth bottleneck, as each top-of-rack switch is linked to the chassis by four parallel 40-Gbps Ethernet links. It adds some latency as compared to a single switch, but the overall latency should be reduced because a larger, distributed switch cuts the number of network hops needed.
The connections between all these components are proprietary, though Juniper doesn't like to describe them as such. "I prefer to use the word unique," said Germanow. He adds that the fabric's roots as a single switch means that it's no more proprietary than any of Juniper's competitors'. "It's just like I can't plug a Cisco line card into someone else's switch."
This is true, but it hints at a wider problem with switches, and one that's going to get a lot worse if the aggregation trend continues. Almost every network manager would love to be able to treat the entire network as a single switch, and every vendor would love to make that happen in a proprietary way.
Making components bigger and further apart seems to run counter to the usual trend in electronics, which is to cram as much functionality into as small a space as possible. However, it's more consistent with the theme of virtualization, centralizing resources into large pools so that they can be reallocated more easily.
Germanow says that the future evolution of QFabric will do both of these: there'll be smaller and cheaper switches where the components eventually come back together again, as well as larger devices that take the place of traditional switch components and continue to scale the network out to even higher capacities.