Cloud server of tomorrow will look little like full-feature server of today

If you have an interest in the architectures that may very well come to dominate the world's most sophisticated data centers, you should take some time to check out an article in EETimes, entitled "Server makers get Goooogled."

The article, by Rick Merrit, describes new technologies being introduced by Rackable and other companies that are strongly influenced by Google's custom server designs over the last several years.

We're talking cool stuff here. As the article notes:

Google has not disclosed details of its motherboard design, but it did release a white paper calling for designs built on 12V-only power supplies. Besides such supplies, Google's design is said to use at least two full servers per board and remove many of the unneeded parts found in many mainstream server motherboards in an effort to shave cost, reduce power consumption and increase reliability.

Bob Warfield of SmoothSpan has a great post on the trend, and why on earth a data center would settle for more of less in server design:

One thing the cloud does, is it will force standardization and penny shaving at the hardware (and software) end. When Amazon, Google, or one of the others is building a big cloud data center, they want utility-grade computing. It has to be dense on MIPS value, meaning it is really compact and cheap for the amount of cpu power delivered. Designs that add 25% to the cost to deliver an extra 10% in power won't cut it. The Cloud will be too concerned about simply delivering more cores and enough memory, disk, and network speed to keep them happy. Closing a deal to build standard hardware for a big cloud vendor will be hugely valuable, and in fact, Rackable started out life building systems for Google.

(Bob also notes that the server world for cloud providers will be a little like the aircraft world for Southwest Airlines--a single standard server architecture utilized routinely throughout the data center for several types of workloads. That's an awesome analog, and one that I've considered many times when guessing at utility computing and cloud computing market trends.)

I've actually wondered for some time about when this would happen--the shift by systems vendors from focusing on enterprise IT to focusing on cloud providers. We are a long way off from that being the dominant model (if, indeed, it ever comes to pass that enterprise IT moves entirely to the cloud). However, Rackable seems to have its sights set on meeting the demands of the Googles and Amazons and Microsofts of the world.

As does IBM, according to GigaOm's Stacy Higgenbotham. She notes that IBM has the iDataPlex product line, which was expressly designed for the cloud:

(iDataPlex servers) have stripped away unnecessary hardware--a move aimed at reducing power-consuming components and saving space. Heat-tolerant processors allow a data center operator to keep air conditioning bills down, saving as much as 4 percent of total energy costs for each degree dropped. So as computing requires more scale, Google's innovations influence other buyers and sellers of technology even as the search giant slows its own data center construction.

Count another one for the Googleplex.

So which vendor should you go with if you are seeking the best cloud experience? Ah, here's the best part. You don't care. Period. As a cloud consumer, which underlying physical hardware comes into play should arguably be a non-issue, and at worst a fleeting thought as you review your cloud options. This is the beauty of the world we are moving to; it's someone else's problem now.

So, those nicely recognizable Dell 2950s or HP DL360 G5s you have stacked in racks in your dev lab will probably be replaced by weird mutant motherboards that couldn't read a USB stick if their life depended on it. You may never know the ultimate hardware architectures that you come to rely on every single day. Yet your livelihood will depend on hundreds or thousands of them.

I, for one, am actually OK with that.

[philcryer]

I think this is a good thing, when I deploy systems I pull things like floppy/cdrom drives and go into the bios and turn off parallel ports, usb, etc. Why power these things if you don't need them? I would love to see systems w/o cdrom drives anymore, installing from a thumbdrive makes more sense, after that all updates are via the web anyway, so there's no need to have cdrom. I have a floater cdrom drive that I hook up just for cd installs, but again, newer systems can boot from usb, so even that's antiquated. So take everything away, have only what is needed to run a server and I'll be happy. I see this as not only the way for cloud computing, but any general purpose production computing - again, why buy and power things you don't need and are not using on your server?

[bonesbautista]

If Google's servers are what's responsible for all of the data hiccups they've had in 2008, then YIKES. I want what they're not using...

[flitcraft33]

Yeah this all groovy and all until you have a problem, call tech support, and find out that you are talking to someone on another sub continent who is reading a PIC on a CRT and has no idea what to do to ACTUALLY help you. Sadly, this is no longer an issue of your TV tuner or home network being off the air, it is for instance, the Excel spreadsheet with the amortization schedule you need for your corporate 1040 or even better, today's payroll for the weekly 941 filing, which alas, I am veddy veddy sohry sir, you will have to recreate, it is now lost, have a nice day....

[meh130]

I always cringe when someone uses Southwest Airline's use of only the Boeing 737 as an illustration of IT standardization. Southwest's use of only Boeing 737s is more an example of a monoculture than a standard. Southwest operates about 540 Boeing 737s. The maintenance, training, and scheduling economies of scale of 540 like types is probably not that much better than operating 270 of type A and 270 of type B. At Southwest's fleet size, you will get efficiencies of scale even if operating multiple types. However, only operating one airplane type means Southwest is risking its entire an $11 billion dollar business on a single airplane type's certificate airworthiness. You will likely find this fact mentioned in the "risks" section of Southwest's 10-K.

This risk was most visible after USAir flight 427 crashed in Pennsylvania, which followed the similar United Airlines flight 585 crash three years earlier. Had the FAA grounded all 737s, Southwest would have been out of business.

That is the risk of a monoculture. The economies of scale of standardization at some point provide diminishing returns. At that point, diversification can provide more strength than standardization. Indeed, it was cramming hundreds of "one size fits all" Pentium 4 based x86 servers into data centers which led to the power and cooling crisis many face.

My recommendations, if your cloud is big enough are the following. It is great to strip down servers to minimize unnecessary power and cooling. Pick a small, medium, and large configuration. Dual source where possible (CPUs, NICs, chipsets, etc.). Move all storage to the network (NFS, iSCSI, or FC). Use hypervisors on all servers, even if it is running one app, as the hypervisor will provide a hardware abstraction layer. Use more than one hypervisor technology (i.e., use some VMware and some Xen) and more than one OS distribution (i.e., use some Red Hat and some SUSE). Software monocultures are riskier than hardware monocultures.

That way, an exploit at the hardware, driver, hypervisor, or OS level will not put your entire enterprise at risk.