HP innovates
I was excited to hear the news recently that HP Networking have been investing in some really cool technology that has the potential to reduce costs, power consumption and enable faster networks.
I was even more interested in this area because before I joined HP, I worked in Nortel's optical division, building a large-scale MEMS based optical switch with an optical mid-plane, and working on remotely controllable optical add/drop multiplexors (ROADMs). At the time the technology was very expensive and suitable only for multi-wavelength (DWDM) long-haul high-capacity systems. (That said… enterprise networking has caught up on the speed side – with our recently announced 40G cards.)
In the subsequent 10 years, HP Labs and HP Networking have worked out how to take similar concepts and dramatically reduce the cost. We have implemented them in a prototype optical backplane, embedded within an HP Networking E8212 modular switch chassis. Charles Clark, one of HP's Distinguished Technologists, demonstrated it at Interop recently.
In chassis-based networking switches, a good proportion of ASIC complexity and power consumption goes into switching packets between interface cards, typically through an electrical cross-bar switch based in ASICs, either on the back of each interface card or in dedicated fabric cards. Data is routed serially at high speed, requiring some pretty complex electronics, high-speed connectors, and careful backplane design to ensure it gets from one card to the next without errors. As the speed of each link increases, the cost, complexity and power consumption of the electronics required to reliably transfer the data between each card increases.
Optics are on a different cost curve
Instead of using electrical transmitters and a costly crossbar switch, HP is exploring the use of low-cost optical transmitters and a fully connected mesh of optical waveguides between each module. This removes the need for the electrical switch, high-speed electrical connectors and a backplane designed for high-speed signals.
An optical backplane such as this (along with the drivers) might take less than one quarter of the power required for a system based on an electrical backplane and crossbar switch.
Whilst fiber optics have been around for a long time, their use has been limited to connecting switches to switches with manually mated connectors on the front of each unit. Using fiber to interconnect interface cards within a system (by connecting a fiber that's mounted on one card to fiber - or other waveguide - on the backplane automatically as you seat the card through a blind-mate connector) has been more of a challenge due to the need for alignment, and cleaning. Instead of using optical fibers to connect between cards, HP has worked out how to connect the outputs from low-cost VCSELs on an interface card to a low-cost injection molded plastic waveguide. The waveguide is plated with a metallic reflective surface, and filled with air.
This is a picture of the actual waveguide assembly complete with a metric ruler for scale
On the receive side, each interface card taps a portion of the light through a passive MEMS-based beam-splitter, and determines if the signal is meant for that card. This allows messages to be sent to multiple cards with just one transmission, making the system very efficient for broadcast and multicast applications.
Whilst the E8200 already has a passive electrical backplane (because the cross-connect is hosted on a dedicated switch card), it still needs careful design to ensure that the electrical signals make it from one card to another without errors. This electrical transmission-line is designed to cope with a certain maximum frequency, which in-turn places a limit on the maximum bitrate between cards.
One of the major benefits of a completely passive optical backplane is that we can start today with interface cards able to communicate using one color of light (wavelength) carrying 120 Gbps (as 12 channels each running at 10Gbps), and then introduce more colors in the future. Though the use of filters, next generation cards could be designed to receive 8, or even 16 times the capacity that a 1st gen card could process initially (and they might even be able to push it to 64, transporting 2Tb of capacity!). With no change to the platform backplane those new, higher capacity interface cards could be introduced while earlier-vintage cards would still function because they would continue ‘listening’ to the original single wavelength.
Demonstrated at Interop
Greg from etherealmind.com was at Interop, and got a chance to take a good look at the system, and described it as "very cool", then went on to say that "HP has demonstrated some serious research work in networking. I’m impressed". Greg also questioned the use of the ProCurve (E-series) chassis rather than the A-Series chassis - but shouldn't read much into the choice of platform for this prototype system. As soon as the technology reaches maturity we'll be using it wherever it makes sense.
Larry Chaffin commented that he thought it was "one of the biggest innovations I saw at the show this year. This will one day greatly speed up switches." Mike Fratto also described it as one of the coolest displays he saw at Interop.
A paper (Low Cost, Injection Molded 120 Gbps Optical Backplane) covering the invention was presented at the National Fiber Optic Engineers Conference in March 2011
When?
After hearing about this technology, the biggest question in my mind for Charles was when this is going to turn into a real product. As I understand it, the transition point will happen when the cost and complexity of making the backplane electronics faster intersects with the decreasing costs of the optical equivalent. Charles thinks this is likely to be sometime in the next 3-to-5 years.
When we introduce this in a product, it will allow us to dramatically ramp performance to support higher-density of faster connections on the front-panel of the system. I hope you'll agree that HP innovation is alive and well in the networking business.
