Change ahead for datacentre networks

Networks in datacentres have consisted of three tiers for a number of years - it may even run into decades. But that looks like it's about to change....

At the top (depending on which way up you hold it) is the access network, which allows you and anyone else with the right permissions to gain networked access to the datacentre's treasures. It connects to the outside world, so here's where you're likely to encounter the heaviest security measures, designed to ensure that only the righteous can get in.

Next is the aggregation layer, which pulls these multiple paths into fewer, higher-speed links, which in turn feed into the core network. While both the top two layers usually run using gigabit Ethernet, data in the core flows at very high speeds, 10 Gbps and above typically, between servers and servers, and servers and storage.

But both Juniper and Cisco have been making noises about a new architecture -- with each vendor's plans being separate from the other's, of course -- that flattens the entire datacentre network into a single layer. They use the word fabric, but frankly I find the word obfuscatory.

Driving this is a need for greater efficiency, and to enable tomorrow's higher speeds and lower latencies. Right now, each time a packet arrives at a switch, it needs to be unpacked to ascertain where it's come from and where it's going. That's a lot of unpacking that occurs millions of times per second -- and in computing terms, it's an expensive operation. This happens at least once per layer so it would save a lot of time and power if packets needed only to be unpacked once, then flowed across a single switch fabric so that any server or other device could connect to any other server or device without passing through more than a single switch.

That's essentially what Juniper is promising, according to company man Trevor Dearing, with whom I had a conversation at Storage Network World last week. He reckons the company has silicon in the labs -- and out in the field with a few selected customers -- that will reduce latency from 50ms to under 10ms. What's more, it'll carry any type of data traffic, including straight Ethernet, Fibre Channel, Convergence Enhanced Ethernet, you name it.

There's a pending IETF standard called TRILL (Transparent Interconnection of Lots of Links, that aims to allow this to work in multi-vendor environments but it's unclear as yet how much effect it will have. There's more about it here.

Other than that, the two datacentre network companies are forging ahead, with Cisco having staked out its claim with FabricPath, which will be based on its high-end Nexus 7000 switches and F-Series line cards.

In terms of technologies, both are working towards what Dearing didn't disagree could be called a layer 1.5 network, sitting just above the physical level, although FabricPath has been described as bringing layer 3 capabilities into layer 2.

You can expect to see the fruits of these labours later this year or early next year, with others promised technology and products including Extreme Networks, Force10 Networks and Brocade.

Networking just became interesting again.