Interest in modular datacentres is growing, fuelled by high-profile endorsements from Microsoft and Google. However, the model raises new management concerns, and efficiency claims may be exaggerated, some say.
Modular, containerised datacentres being sold by vendors such as IBM, Sun and Rackable Systems fit storage and hundreds of servers — sometimes thousands — into one large shipping container with its own cooling system. Microsoft, using Rackable containers, is building a datacentre outside Chicago with more than 150 containerised datacentres, each holding 1,000 to 2,000 servers. Google, not to be outdone, secured a patent last year for a modular datacentre that includes "an intermodal shipping container and computing systems mounted within the container".
To hear some people tell it, containerised datacentres are far easier to set up than traditional ones, and are also easier to manage and more power-efficient. It should also be easier to secure permits, depending on local building regulations. Who wouldn't want one?
If a business has a choice between buying a shipping container full of servers, and building a datacentre from the ground up, it's a no-brainer, says Geoffrey Noer, a vice president at Rackable, which sells the ICE Cube Modular Data Center.
"We don't believe there's a good reason to go the traditional route the vast majority of the time," he says.
But that is not the consensus view by any stretch of the imagination. Claims about efficiency are over-rated, according to some observers.
Even IBM, which offers a portable modular datacentre and calls the container part of its green strategy, says the same efficiency can be achieved within the four walls of a normal building.
IBM touts a modular approach to datacentre construction, taking advantage of standardised designs and predefined components, but that doesn't have to be in a container. "We're a huge supporter of modular. We're a limited supporter of container-based data centers," says Steve Sams, vice president of IBM Global Technology Services.
Containers are efficient because they pack lots of servers into a small space, and use standardised designs with modular components, he says. But it's possible to deploy storage and servers with the same level of density inside a building, he notes.
Container vendors often tout 40%- 80% savings on cooling costs. But according to Sams, "in almost all cases they're comparing a highly dense [container] to a low-density [traditional datacentre]".
Containers also eliminate one scalability advantage related to cooling found in traditional datacentres, according to Sams. Just as it's more efficient to cool an apartment complex with 100 living units than it is to cool 100 separate houses, it's more cost-effective to cool a huge datacentre than many small ones, he says. Air conditioning systems for containerised datacentres are locked inside, just like the servers and storage, making true scalability impossible to achieve, he notes.
Gartner analyst Rakesh Kumar says it will take a bit of creative marketing for vendors to convince customers that containers are inherently more efficient than regular datacentres. Gartner is still analysing the data, but as of now Kumar says, "I don't think energy consumption will necessarily be an advantage."
That doesn't mean there aren't any advantages, however. A container can be up and running within two or three months, eliminating lengthy building and permitting times. But if you need an instant boost in capacity, why not just go to a hosting provider, Kumar asks.
"We don't think it's going to become a mainstream solution," he says. "We're struggling to find real benefits."
Kumar sees the containers being more suited to internet-based, "hyper-scale" companies such as Google, Amazon and Microsoft. Containerised datacentres offer scalability in big chunks, if you're willing to buy more containers. But they don't offer scalability inside each container, once it has been filled, he says.
Container vendors tout various benefits, of course. Each container is almost fully self-contained, Rackable's Noer says. Chilled water, power and networking are the only things from the outside world that must be connected to each one, he says. Rackable containers, which can be fitted with as many as 22,400 processing cores in 2,800 servers, are water-tight, fitted with locks, alarms and LoJack-like tracking units. Sun's modular datacentre can survive an earthquake —the company made sure of that by testing it on one of the world's largest shake tables at the University of California in San Diego.
A fully-equipped Rackable ICE Cube costs several million dollars, mostly for the servers themselves, Noer says. The container pays for itself with lower electricity costs due to an innovative Rackable design that maximises server density, Noer says.
But it's still too early to tell whether containerised datacentres are the way of the future. "We're just at the cusp of broad adoption," Noer says.
Potential uses for containers include disaster recovery, remote locations such as military bases, or big IT hosting companies that would prefer not to build brick-and-mortar data centers, Kumar says.
A TV crew that follows sporting events may want a mobile data center, says Robert Bunger, director of business development for American Power Conversion. APC doesn't sell its portable datacentre, but in 2004 it built one into a tractor-trailer as a proof-of-concept. It was resilient.
But APC isn't seeing much demand, except in limited cases. For example, a business that needs an immediate capacity upgrade but is also planning to move its datacentre in a year might want a container because it would be easier to move than individual servers and storage boxes.
The University of California at San Diego bought two of Sun's modular datacentres. One goal is to contain the cost of storing and processing rapidly increasing amounts of data, says Tom DeFanti, principal investigator of the university's GreenLight energy efficiency research project. But it will take time to see whether the container approach is more efficient. "The whole idea is to create an experiment to see if we can get more work per watts," DeFanti says.
The modular datacentre is not as convenient to maintain as a regular computer room, because there is so little space to maneuver inside, he says. But "It seems to me to be an extremely well-designed and thought-out system," DeFanti says. "It gives us a way of dealing with the exploding amount of scientific computing that we need to do."
Before purchasing a containerised datacentre, organisations should consider several issues related to their manageability and usefulness. Vendors often want you to fill the containers with only their servers, Kumar notes. Besides limiting flexibility at the time of purchase, this raises the question of what happens when those servers reach end-of-life. Will you need the vendor to rip out the servers and put new ones in, once again limiting your choice of technology?
"At the moment, most vendors will fill their containers only with their servers," Kumar says.
IBM, however, says it uses industry-standard racks in its portable datacentre, allowing customers to buy whatever technology they like.
DeFanti says Sun's modular datacentre allows him the flexibility to buy a heterogeneous mix of servers and storage. Rackable, though, steers customers toward either its own servers or IBM BladeCenter machines through a partnership with IBM.
"I think vendors are learning that people want more flexibility," DeFanti says.
Another consideration is failover capabilities, says Lee Kirby, who provides site assessments, data center designs and other services as the general manager of Lee Technologies.
If one container goes down, its work must be transferred to another. Server virtualisation will help provide this failover capability, and also make it easier to manage distributed containerised datacentres — an important consideration for customers who want to distribute computing power and have it reside as close to users as possible, Kirby says.
"I think it is key that the combination of virtualisation and distributed infrastructure produce a container that can be out of service without impacting the application as a whole," Kirby says.