Alcatel-Lucent has introduced a data center switch fabric product strategy that promises the ability to build a two-tiered network with just two core data center chasses and 14,400 server-facing 10 Gigabit Ethernet (GbE) ports for 5 microsecond end-to-end latency.
Alcatel-Lucent claims it can achieve this network with its new Alcatel-Lucent Enterprise Mesh technology, which interconnects racks of servers through a new top-of-rack switch, the OmniVista 6900, managing them as "pods" of six switches via multiple 40 GbE uplink ports and Multi-Chassis Link Aggregation Group (MC-LAG or MLAG) technology. These six-switch pods then interconnect with other pods to form a "mesh." Each mesh contains five pods and a total of 30 top-of-rack switches. Finally, each mesh uplinks via 40 GbE into a pair of OmniVista 10000 chassis switches.
Everyone has a data center switch fabric these days
Alcatel-Lucent is only the latest networking vendor to pitch a data center switch fabric to enterprises. Cisco has introduced FabricPath and its Nexus switches, Juniper has QFabric, Brocade has its Brocade One strategy and its VDX switches and Avaya has its Virtual Enterprise Network Architecture (VENA) strategy born out of its Nortel Networks acquisition. Startup Arista Networks is also offering a solution built on high-performance switches and MC-LAG technology, although it has eschewed describing its approach as a data center switch fabric. HP is the only major networking vendor that hasn't introduced a data center switch fabric concept, although that may change at Interop next month.
"They all eventually solve the same problem: Make the network in the data center more efficient," said Zeus Kerravala, senior vice president and distinguished research fellow at Yankee Group. "But right now it's pretty wide open as to who is going to win this. In a lot of cases, these vendors don't have a product out yet."
Alcatel-Lucent's data center switch fabric awaits 40 GbE support for full performance
Like most of these vendors’ products today, Alcatel Lucent's Enterprise Mesh architecture is more vision than reality. Until the vendor delivers support for 40 GbE, the total end-to-end server latency of 5 microseconds lives only on marketing slides. Customers can build out a mesh using 10 GbE uplinks on the OmniSwitch 6900 rather than 40 GbE, but the total latency will be around 22 microseconds.
"Customers are definitely looking toward and inquiring about 40 GbE capabilities," said Mike Spanbauer, senior analyst with Current Analysis. "When you have this many 10 GbE ingresses, how are you going to get it off the box fast enough to meet applications' needs? You need to have a switching architecture that is easily cabled because you don't have to have trunks and trunks of 10 GbE."
The new OmniSwitch 6900, available this quarter, features 64 10 GbE ports, 1.28 Tbps total throughput and capacity for 128,000 MAC addresses. It will support six 40 GbE uplinks in the second half of 2011. In a fully deployed mesh, five of those 40 GbE uplinks will interconnect with other switches in the local "pod," and the sixth uplink will connect into the larger mesh.
The switch also comes in 20- and 40-port versions for different rack configurations. Alcatel-Lucent claims the switch consumes 3.5 watts of power per 10 GbE port.
The OmniSwitch 6900 is comparable to Juniper Networks' QFX 3500 and Cisco Systems' new Nexus 3000, Spanbauer said. He said all three switches rely on merchant silicon from Broadcom's Trident chip family. All three are high-performance, low-latency top-of-rack switches that will support 40 gigabit bandwidth eventually. In the meantime, customers can experiment with what's available by building pods and meshes of the OmniVista 6900 switches using 10 GbE uplinks, much like Brocade customers can build small data center fabrics with the VDX 6720 top-of-rack switch while awaiting delivery of the company's larger VDX chassis product.
"I don't think companies are all of a sudden going to say, 'I'm going to build a fabric' and rip out all the other stuff," Kerravala said. "It's likely that they will put up a fabric in one part of their data center, attach a new service to it and grow it from there. That's one advantage of [Alcatel-Lucent's] six-switch pod. You can get into that with relatively low costs and connect them together and have a mini-fabric. Then you can decide if that's what you want. You can then buy a core switch and run it as a traditional fabric—as if there is such a thing as a traditional fabric."
Alcatel-Lucent customers can also use MC-LAG to eliminate spanning tree protocol in its data center fabric and to lay the groundwork for storage and data convergence. In the future, the company will add support for lossless Ethernet and Shortest Path Bridging, according to Cliff Grossner, director of product marketing for Alcatel-Lucent Enterprise.
Data center switch fabric relies on application and virtualization intelligence, control
Beyond super-low latency, Alcatel-Lucent's strategy also includes tighter management for a very complex environment. With so many interconnections of top-of-rack switches, the Alcatel-Lucent data center switch fabric emphasizes server-to-server traffic and minimizes north-south traffic to the twin core switches in the design. In such a design, the network needs more visibility and control into the applications and virtual machines. So Alcatel-Lucent has introduced the Application Fluent Network. The company relies on a technology called Virtual Network Profile (vNP), which recognizes individual applications. vNP automatically optimizes the network for the individual application, adjusting QoS and security profiles based on policies set by the networking team.
Alcatel-Lucent has also introduced a new module for its OmniVista 2500 network management platform, Virtual Machine Management (VMM). VMM integrates with all major hypervisor products and allows network managers more control and visibility into virtual machine (VM) behavior. It can automatically direct a VM to migrate from one pod to another in order to make server-to-server traffic more efficient.
These technologies give Alcatel-Lucent's data center switch fabric "the ability to set policies in the environment which automatically bind to virtual machines and virtual machine location as well as provide automatic provisioning of applications within the environment," Spanbauer said. "You preset this and it will self-apply. The network itself can actually request a virtual machine move to optimize an application or to minimize latency for certain latency-sensitive applications."
How will network engineers choose from switch fabric options?
It's tough to know whether better management and visibility or a pod approach will sway customers toward Alcatel-Lucent, a vendor with limited visibility in North America. With each new data center switch fabric that hits the market, network engineers are faced with an increasingly complex range of choices.
"Every vendor today looks somewhat different," Kerravala said. "Prior to this pivot point, the real innovation [in networking] was either making something cheaper or denser and a little bit faster. But essentially everything looked the same. This is the first time in a long time we've seen real differentiation in the network space."
The old approach of looking at vendors on a switch-by-switch basis will no longer apply for network engineers, said Kerravala. For instance, how can you compare Juniper's QFabric Interconnect chassis with competing products when it's not a true Ethernet switch?
"Evaluating these fabrics can't be done on the same criteria you used to evaluate a single switch," Kerravala said. "You can't say, 'Well, it's got 48 ports and this much switching capacity.' All of these switches are designed to run in a complete network so you need to think about what network provides the best server-to-server latency time. What's the aggregate switch capacity of the whole network? Not a single box."
Let us know what you think about the story; email Shamus McGillicuddy, News Editor.