Google: We Want a New 100G MSA
opticalreflection.com
By PAULINE | Published: SEPTEMBER 21, 2010
Google caused a bit of a stir at the Market Focus Session during the ECOC exhibition here in Turin.
The internet search giant wants to use 100Gigabit Ethernet (100GbE) inside its massive data centres, but it can’t find an optical module that meets its requirements. So it’s asking the optical components companies to develop a new type of optical module – a request that risks dividing the 100GbE market.
“We have a significant demand for bandwidth that we can’t meet simply by scaling 10Gbps ports,” said Bikash Koley, senior network architect at Google. “We’re ready to deploy in volume. It’s just that we don’t have the right form factor to deploy today.”
The details of 100GbE have been thrashed out by the IEEE 802.3ba Working Group over the past three years. And as is traditional in the world of Ethernet, they specified several varieties of 100GbE to suit different applications. 100GBase-SR10 is designed to send signals using 10 parallel channels over 125m, far enough to reach across the room inside most data centres, while 100GBase-LR4 uses four wavelengths multiplexed onto a single fibre, which can reach distances up to 10km for campus type applications.
But Google’s massive data centres probably aren’t typical in their requirements.
Koley dismisses 100GBase-SR10 in short order. “100m is not sufficient; we can’t even go room to room with that,” he pointed out, adding: “Ribbon fibres are hard to deploy, hard to manage, hard to terminate and hard to connect. We don’t like them.”
100GBase-LR4 will go longer distances, but Google says the technology hasn’t reached “utility” status; in other words, it needs to be easier to deploy, cheaper and to consume less power than the technology it is replacing. Right now that isn’t the case, he contends. The CFP form factor that has been developed for 100GbE may be pluggable, but it is physically quite large. And not only are the modules more expensive, they actually consume more power rather than less.
Currently available 100GBase-LR4 optical modules guzzle 20W, which compares rather unfavourably with ten 10 Gbps SFP+ modules each consuming less than 1W. When you’ve got as many interfaces as Google has in its data centres, a doubling in power consumption adds up to a very large, unacceptable number.
From Google’s point of view the part is overdesigned. The company isn’t willing to pay the price or power premium to get to 10 km when 2km would suffice. Nor does it want the interface timing function that is included in the module to support Sonet jitter requirements; the power to provide that timing function is just power wasted when the function won’t get used.
Finally, there’s what Koley calls the “25G gearbox problem”. The electrical wiring on printed circuit boards in servers and switches runs at 10 Gbps today, whereas the four lasers and detectors inside the module are driven by 25Gbps signals. The “gearbox” is a serialize-deserializer IC that is necessary to convert from one data rate to the other.
The data centre industry is moving towards 25 Gbps interfaces for ICs, which would eliminate the need for the conversion, but the intellectual property for those chips is only just being developed now and probably won’t reach the market for another four years, says Koley. In the meantime, the modules are stuck with the additional part, with the associated cost and power consumption.
Google has a pretty specific idea about what do instead: it’s asking components vendors to develop a completely new 100GbE multisource agreement (MSA) for a low cost, low power, single-mode interface with 2 km reach that can also be non-retimed. The firm reckons that ten-channel laser arrays that have been developed for other applications could be adapted to meet this brief in a short time frame.
“We are inviting manufacturers to put together an MSA because this is a real need,” Koley said.
How have the components vendors responded?
Finisar’s engineering director Chris Cole has clearly had time to think this over because his presentation, which followed Koley’s, warned against taking commandments from “Mount Google”.
“Investment fragmentation into competing 4×25 and 10×10 standards is the worst possible scenario or the future of 100G, and will delay the development of the market,” he said. Standards are meant to be exactly that: a single solution to a problem that allows everyone to benefit from economies of scale.
“Some people are flippantly dismissing the three-year effort by the IEEE and what I can tell you is that I resent that,” he added, pointing out that the technology alternatives were exhaustively studied in the IEEE Working Group in an open and democratic process. When it came to a vote, 4 x 25 Gbps was explicitly supported and 10 x 10 Gbps was explicitly rejected by participants representing an overwhelming industry majority.
Finisar accepts that the current 100GBase-LR4 module is expensive – more than twice the potential cost of a ten-channel module – but says that prices will come down next year and will continue to fall because the 4×25 Gbps configuration is fundamentally a lower cost form factor.
The optics cost today is dominated by the cost of using discrete EML lasers and the fact that the “gearbox” IC hasn’t reached mainstream production, but is coming off multi-project wafers. Next year Finisar plans to introduce a module containing an integrated four-laser array, which it expects to come in at around 60% of the cost of the ten-channel chip.
Cole also pointed out that the supply of ten-channel laser arrays is risky because the technology isn’t mature, and is only available from a single source. In contrast, 4×25 Gbps modules are already available from three vendors.
But despite the disagreement, the talk ended on a conciliatory note. “We hear your needs”, Cole told Google. “100GBase-LR4 in a QSFP form factor might be the solution you are looking for.” |
