Intel Considering Portable Data Centers

miller60 writes "Intel has become the latest major tech company to express interest in using portable data centers to transform IT infrastructure. Intel says an approach using a "data center in a box" could be 30 to 50 percent cheaper than the current cost of building a data center. "The difference is so great that with this solution, brick-and-mortar data centers may become a thing of the past," an Intel exec writes. Sun and Rackable have introduced portable data centers, while Google has a patent for one and Microsoft has explored the concept. But for all the enthusiasm for data centers in shipping containers, there are few real-world deployments, which raises the question: are portable data centers just fun to speculate about, or can they be a practical solution for the current data center expansion challenges?"

Re:Why it probably won't work

[Feyr]

good points, and there's also the maintenance and upgrades to consider, unless you're google and you just replace the rack when more than a certain % is defective. for the majority of places, clustering is the exception, not the norm and you just cant leave 70% of your rack full of defective or outdated crap

consider minor faults too. do you replace the whole rack because a network cable went bad? i don't think so, and i don't want to be the one crawling around that shipping container stringing cat5

Military

[SirKron]

The military already uses these. The Marines uses them to bring their network onto a ship during transit and then into a tent when deployed.

Like Prefab Houses

[Doc Ruby]

Prefab houses are an increasingly popular method for home construction. They're not really "portable", except when they're delivered from the factory to the "installation site". They're not interesting because of their containers, but because of the economics and other efficiencies in delivering and installing them.

Instead of the house builders building each house as a completely custom job, in an unfamiliar site, in all kinds of weather, with only the tools and materials they bring to some residential area, they've got full control at the factory. They don't have to ship all the excess materials that they used to have to ship back out as garbage. They can keep a pipeline filled with houses they're building, and deliver them very shortly after they're ordered, even quicker than they actually build them. And since so much is standardized, they can mass produce them and otherwise get scale economies that reduce costs. Since they aren't inventing a new, complex device with every home a new, arbitrary blueprint, they are skilled in more than their tools and materials, but rather skilled in producing that exact house, with solved problems presenting higher quality homes quicker.

All that is also true of datacenters. The weather doesn't present so much of a problem avoided, because the datacenter is usually installed in an existing building. But all the rest of the efficiencies are in effect. So datacenters can be cheaper, better, and deployed quicker. This trend makes a lot of sense.

Re:Why it probably won't work

[TheLink]

"You need to be able to access all the equipment"

Why? If you're something like Google, I bet you could just RMA the containers with faulty stuff back and get new/refurbished ones already configured to your specs - all you need is net boot them for automated install. AFAIK Google don't fix servers once they fail or even take them out of the rack, they just have someone go about once in a while to take em out (like "garbage collecting" instead of "malloc/free").

So for the big guys it'll be a bit like buying a prebuilt PC, only it's the size of a container.

Re:Why it probably won't work

[dokebi]

Google isn't doing that just because they have lots of money. No, it's actually cheaper to run things that way. And now with VM's running on clusters, the health of individual machines really doesn't matter anymore.

So, when do you think a Redundant Array of Inexpensive Datacenters will become a reality? Psst. It'll be sooner than you think.

Re:It has to be more expensive

[Kadin2048]

Rule #1 in technology, anything portable is more expensive than if it were not portable. If its so cheap to use a crate, why not just put the stuff in the crate in a warehouse instead, bypassing the crate and all of the work and design involved with shoving and fitting the stuff in the crate?

Not really applicable here. The equipment is the same either way. It's not like buying a laptop versus a desktop, where one is carefully (and expensively) optimized and the other one isn't. The same pizzaboxes/blades are going in the racks either way, whether it's in a traditional datacenter or in a cargo container.

The advantage is more on the installation and infrastructure end. Think of it more as "mobile homes" versus "traditional houses." With a regular house, you have to get the plumber, electrician, HVAC guy, carpenters, etc. to your site. For a mobile home or trailer, you keep all those people in one place, and they build houses over and over and over, on an assembly line. And as a result, "manufactured homes" are a lot cheaper than regular ones.

I think that's the model that you want to apply to datacenters: get rid of all the on-site installation and configuration, all the raised flooring and cabling; just have a team of people in a factory somewhere, installing and wiring all the servers into the containers, over and over. Then you just haul the container to the customer's site and plug it in. (In fact, since it's in a shipping container already, there's no reason why you do this in a place where labor is expensive; you might as well assemble them in some third-world country somewhere; it would almost assuredly be worth the small cost for sea freight -- most of a container's transportation costs are in the last few hundred miles anyway.)

The problem is mainly a chicken-and-egg one; in order to make "datacenters in a box" cheaper than traditional ones, you need to get an economy of scale going. You need to have an assembly line churning them out. If you don't have that, you're just taking the expense of a traditional data center and then adding a bunch of containerization and transportation costs to it.

It might take a very long time to catch on, because there's such an investment in traditional datacenters right now, but if I worked doing datacenter server installations, it's probably something I'd be a little concerned about. Unlike with 'manufactured homes' and regular houses, there isn't much social stigma over having your web site served from a trailer.

Re:Why it probably won't work

[Kadin2048]

I think the short answer is that you don't. I've seen the photos of Sun's boxes, and while the racks do pull out to let you get to the equipment if you need to, I think you basically just view each server in the rack as a small part of a bigger assembly (the box itself), and if something goes faulty in a single server, you move its workload to another machine and just turn it off and leave it there, essentially entombed in the rack. Maybe they'll be some way of easily swapping out machines, or maybe it'll just be easier to leave them there until the entire container's worth of machines are obsolete, and then just dispose of the whole thing and get a new box hauled in. (Or send it back to somewhere for refurbishment, where they can strip it down completely, pull out all the machines, repair and replace, and then bring in a new one.)

We think of rack space as being precious because of the way traditional datacenters are built and designed; I'm not sure that would still be true if you had a warehouse or parking lot full of crates (especially if they're stacked 3 or 4 high) instead. If you never unseal the box, rack space isn't a concern. Heck, if you have a football field of stacked containers, you might not even want to mess around with getting a dead one out of a stack if it died completely. Just leave it there until you have some major maintenance scheduled and it's convenient to remove it.

This is getting into business models rather than the technology itself, but I could imagine a company selling or leasing boxes with a certain number of actual processing nodes and a number of hot spares, and a contract to replace the container if more than x number of nodes failed during the box's service life (5 years or so). Companies could buy them, plug them in, and basically forget about them, like the old stories about IBM mainframes. If enough units in the box failed so that it was close to running out of hot spares, then it could phone home for a replacement. As long as enough hot spares were provided so that you didn't need to do this often, it might be fairly economical.