Some of these suggestions are bold, but it’s hard to see how Dell can continue to compete against IBM, HP and Cisco on one end and cloud services on the other without taking some risks. I actually think Dell is well positioned to be a big part of the future of IT if it just embraces a future that’s all about software and next-generation applications. Hardware, well, it’s largely just a necessary evil on which increasingly intelligent and virtualized software needs to run.

1. One word (occasionally hyphenated): Webscale.

Here’s an interesting fact: According to IDC, while the rest of the big four server vendors lost both market share and revenue in the third quarter of 2012, Dell gained in both. Why? Because it’s selling webscale servers like crazy — thousands at a time to web companies such as eBay, cloud providers such as Microsoft and even high-performance computing shops. In fact, Dell has shipped more than 1 million of these no-frills, scale-out boxes since it began selling them in 2007.

It seems only logical that Dell should keep riding this horse on the hardware side (including by continuing its close relationship with the Open Compute Project), but there’s a real opportunity in software, too. Although companies like Facebook and Google have the engineers to build their own tools for managing large server environments — both in terms of operations and power consumption — many smaller and more-mainstream companies don’t. If it’s already selling the servers, Dell might as well be selling the software that ensures they’re easy to manage and troubleshoot, too.

2. Make a big bet on big data.

If Dell wants to get a piece of a big data market that’s going to be worth tens of billions in the next few years, it will have to do more than just sell servers. Dell needs a database or platform like Hadoop that it can deliver as a service on top of its commodity hardware — much like what Oracle, EMC, IBM and others are trying, only with a lower price tag and a focus on scaling out versus scaling up.

I actually suggested Dell make a big data investment (and cloud and systems-management investments) two years ago, and although a lot of potentially good fits have been bought in the meantime, there’s still a variety of options available. The Hadoop distribution vendors are probably out of its price range (so closer partnerships with Cloudera, Hortonworks and/or MapR should be in order), but there are whole markets full of analytic, NoSQL and NewSQL databases just waiting to be snatched up.

3. Buy up some real cloud knowledge.

Sorry to burst any bubbles, but it’s hard to see how Dell’s proposed Microsoft, VMware and OpenStack-based public cloud offerings will amount to much unless they’re part of a division with an ingrained cloud mentality. The company is competing in this space against not only Amazon Web Services, but also the companies (and Rackspace) who have developed and are selling the very platforms Dell is looking to resell. And the debacle that is HP’s cloud business should highlight the perils of trying to build something from scratch when operating cloud platforms isn’t in your corporate DNA.

Still, I think OpenStack is probably Dell’s best bet (if only because it has already made its bed with the open source platform). Like Rackspace with Anso Labs (and unlike the DIY approach by HP), Dell could acquire some OpenStack talent to help it build quality cloud infrastructure offerings — both public and private — and services. And even outside of OpenStack, there are still a few startups that have real institutional knowledge of building and running cloud infrastructure, but that might not come with a huge price tag.

4. Push the envelope with ARM.

One of Dell’s ARM server designs

Dell’s focus on alternative server architectures shouldn’t stop with its webscale business, but should color its entire hardware strategy. And if it wants to be the server maker for next-generation applications, doubling down on its ARM investment is probably a good thing to do. Dell has already designed a couple of ARM servers, but they’ve generally been reserved for webscale and research customers, as well as for the Apache Software Foundation.

However, as even Dell has acknowledged, many big data and other applications that rely on distributed infrastructure and parallel processing don’t require brawny x86 chips. They just needs lots of cores, which is something ARM’s server designs can deliver at a fraction of the energy footprint. Assuming software vendors are on board — and a growing number appear to be — Dell could actually distinguish itself by being the only big server vendor delivering tomorrow’s applications on gear actually designed for their unique needs.

5. Tell Microsoft to suck it.

Or, rather, gently ask its new creditor and likely board member to take a back seat and hope the rest of the board agrees. Because if there’s one thing Dell can’t afford to do, it’s get locked up in the same last-millenium relationship that has both companies reeling today. Yes, there might — and maybe should — be some deals between the two companies to deliver Microsoft software on Dell hardware, but in an era of open source, cloud computing, VMware and Apple, there’s no real future for Dell in pushing Windows anything down its customers’ throats.

Besides, I don’t think Microsoft actually needs to push its agenda in order for the Dell investment to pay off (and it’s not just because a successful Dell is the only way to make Microsoft’s investment pay off). Microsoft, too, is struggling to find its direction in a world increasingly dominated by alternative operating systems (mobile, desktop and server) and it should realize there’s precious little Dell hardware can do to get people to buy Windows.

If anything, Microsoft should view Dell as a great server vendor to power its Windows Azure and Office 365 cloud offerings and provide the physical medium for some next-generation enterprise software.

Feature image courtesy of Shutterstock user Dusit.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Infrastructure Q1: IaaS Comes Down to Earth; Big Data Takes Flight
• The promise of SDNs in the enterprise
• Cloud computing infrastructure: 2012 and beyond

Pinterest Operations Engineer Ryan Park had the stage to himself for a session on Wednesday, while Flipboard Chief Architect Greg Scallan and Yelp Engineering Manager Jim Blomo teamed up with Kleiner Perkins Caufield Byers Partner Ray Bradford to form a trifecta of wisdom on Thursday.

Know — and measure — your costs

Flipboard’s Scallan had a paradoxical lesson for the audience when it comes to managing cloud-based infrastructure: Embrace the cloud, but be afraid of the cloud. Yes, it’s flexible and affordable if done right, but all it takes is poor planning or a handful of servers left running ad infinitum, and the costs can begin to grow out of control. That’s why Flipboard assigns members of its engineering team the title of “chief miser,” which means they’re the ones who decide that applications are using the right resources and using them wisely.

Thanks to a variety of practices, including its miserly ways, Scallan said Flipboard is now running about 900 instances at any given time. That’s down from a peak of about 1,500.

Some stats on Flipboard’s AWS usage

One way to help ensure this sort lean operation is to understand your business inputs and outputs, Kleiner Perkins’s Bradford explained. He suggests companies ask, for example, what it costs them to serve a free user on their platform and how does that change with scale or affect the experience they can offer premium users. Pick metrics that really matter, he said (e.g., infrastructure cost per user per month) and then consider how long your current  architecture can sustain that cost before it’s time to retool.

The secret weapon: Source your instances wisely

Pinterest, Yelp and Flipboard all swear by AWS’s pre-paid Reserved Instances in order to save money over the long haul. In fact, Flipboard’s Scallan said, the e-reading startup sees cost savings of about 80 percent over three years by using heavy-duty Reserved Instances instead of on-demand instances for its base workloads, and the break-even point might be only eight or nine months. Pinterest’s Park cited savings of about 70 percent over three years using them.

The trick is queuing another job to take up the waste.

Yelp’s Blomo said his company is a heavy Elastic MapReduce (EMR) user, peaking at more than 350 Elastic MapReduce instances when many developers run their Hadoop jobs simultaneously or when it’s doing nightly analysis of its log files. In order to keep costs in check, Yelp uses Reserved Instances whenever possible to save on hourly bills and has implemented a job-flow pooling system to keep Hadoop jobs running continuously as resources become available. This helps avoid the situation where a job completes in 61 minutes, for example, thus triggering the charge for a full hour of resources even though it only used a minute worth of the second hour.

In order to best gauge when it should use what type instance, Yelp created a tool called EMRio that analyzes past usage to determine what resources are the most-efficient choice for any given job.

The results of EMRio

When it comes to optimizing costs on AWS, though, Pinterest appears to have it all figured out — even how to make use of the somewhat tricky Spot Instances that are priced based on demand and can be terminated without notice if the market price outgrows a user’s bid. Park explained how Pinterest uses the heck out of Reserved Instances and created its own auto-scaling “watchdog” service that decides whether to use Spot Instances or on-demand instances when more resources are required.

Ryan Park dropping knowledge — and graphs

Although Spot Instance prices occasionally spike through the roof, Park’s experience is that they typically remain stable and can result in “massive” savings if you know how to use them effectively. Using Spot Instances to power Pinterest’s approximately 80 front-end servers costs only about $20 per hour, he said. All told, Pinterest has reduced its daily computing bill to about $440 from about $1,200.

All this being said, though, Park, Blomo and Scallan all acknowledged that the flexibility of being able to mix on-demand, reserved and spot servers might not be all it’s cracked up to be if you don’t understand how they all work. Reserved Instances are inflexible in terms of size and region once you reserve them, and Spot Instances must be used wisely for jobs or applications that can handle their easy come, easy go nature. And now there’s even more to consider because Reserved Instances can be resold via AWS’s spot marketplace.

“It gets a little tricky,” Blomo said.

Pick your challenges

Although decisions such database type and structure are largely architectural, there might be elements of cost efficiency at play, as well. Maybe Kleiner Perkins’s Bradford put it best while leading off the session with Scallan and Blomo. Bradford presented a slide containing a simple quote from Instagram Founder Mike Krieger: “Your users around the world don’t care that you wrote your own database.” Sometimes, Bradford added, it might be best to use what works — maybe even a managed service — rather than whatever’s trending highest on Hacker News.

Pinterest’s Park expressed a similar sentiment during his session, citing a lesson his team learned about trying out too many new databases. The site used to use MongoDB, Cassandra, Redis and other databases simultaneously, but learning all the new technologies and managing them became burdensome. Now, he said, Pinterest uses good, old-fashioned MySQL (granted, it sharded MySQL 4,000 times) and memcached — as well as Redis — because they have strong communities and new engineers are more likely to know how to work with them.

After explaining EMRio and some other custom-built Hadoop tools to the crowd, Yelp’s Blomo noted that companies should carefully consider whether the time and money it takes to build stuff will actually result in commensurate savings once those tools or systems are in production. That can require some tough balancing of criteria such as cost, performance, flexibility and user experience.

But it’s important to use human resources wisely. As Bradford said during his presentation, “There’s no free lunch when it comes to developer time.”

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Takeaways from the second quarter in cloud and data
• How direct-access solutions can speed up cloud adoption
• Understanding and managing the cost of the cloud

As with most of its changes to Hadoop over the years — including the recently unveiled AvatarNode — Corona came to be because Hadoop simply wasn’t designed to handle Facebook’s scale or its broad usage of the platform. What kind of scale are we talking about? According to Facebook engineers Avery Ching, Ravi Murthy, Dmytro Molkov,‎ Ramkumar Vadali, and Paul Yang in a blog post detailing Corona on Thursday, the company’s largest cluster is more than 100 petabytes; it runs more than 60,000 Hive queries a day; and its data warehouse has grown 2,500x in four years.

Further, Ching and company note — echoing something Facebook VP of Infrastructure Engineering Jay Parikh told me in September when discussing the future of big data startups — Hadoop is responsible for a lot of how Facebook runs both its platform and its business:

Almost every team at Facebook depends on our custom-built data infrastructure for warehousing and analytics, with roughly 1,000 people across the company — including both technical and non-technical personnel — using these technologies every day. Over half a petabyte of new data arrives in the warehouse every 24 hours, and ad-hoc queries, data pipelines, and custom MapReduce jobs process this raw data around the clock to generate more meaningful features and aggregations.

So, what is Corona?

In a nutshell, Corona represents a new system for scheduling Hadoop jobs that makes better use of a cluster’s resources and also makes it more amenable to multitenant environments like the one Facebook operates. Ching et al explain the problems and the solution in some detail, but the short explanation is that Hadoop’s JobTracker node is responsible for both cluster management and job-scheduling, but has a hard time keeping up with both tasks as clusters grow and the number of jobs sent to them increase.

Further, job-scheduling in Hadoop involves an inherent delay, which is problematic for small jobs that need fast results. And a fixed configuration of “map” and “reduce” slots means Hadoop clusters run inefficiently when jobs don’t fit into the remaining slots or when they’re not MapReduce jobs at all.

Corona resolves some of these problems by creating individual job trackers for each job and a cluster manager focused solely on tracking nodes and the amount of available resources. Thanks to this simplified architecture and a few other changes, the latency to get a job started is reduced and the cluster manager can make fast scheduling decisions because it’s not also responsible for tracking the progress of running jobs. Corona also incorporates a feature that divvies a cluster into resource pools to ensure every group within the company gets its fair share of resources.

The results have lived up to expectations since Corona went into full production in mid-2012: the average time to refill idle resources improved by 17 percent; resource utilization over regular MapReduce improved to 95 percent from 70 percent (in a simulation cluster); resource unfairness dropped to 3.6 percent with Corona versus 14.3 percent with traditional MapReduce; and latency on a test job Facebook runs every four minutes has been

Despite the hard work put into building and deploying Corona, though, the project still was a way to go. One of the biggest improvements currently being developed is to enable resource management based on CPU, memory and other job requirements rather than just the number of “map” and “reduce” slots needed. This will open Corona up to running non-MapReduce jobs, therefore making a Hadoop cluster more of a general-purpose parallel computing cluster.

Facebook is also trying to incorporate online upgrades, which would mean a cluster doesn’t have to come down every time part of the management layer undergoes an update.

Why Facebook sometimes must re-invent the wheel

Anyone deeply familiar with the Hadoop space might be thinking that a lot of what Facebook has done with Corona sounds familiar — and that’s because it kind of is. The Apache YARN project that has been integrated into the latest version of Apache Hadoop similarly splits the JobTracker into separate cluster-management and job-tracking components, and already allows for non-MapReduce workloads. Further, there is a whole class of commercial and open source cluster-management tools that have their own solutions to the problems Corona tries to solve, including Apache Mesos, which is Twitter’s tool of choice.
However, anyone who’s familiar with Facebook knows the company isn’t likely to buy software from anyone. It also has reached a point of customization with its Hadoop environment where even open-source projects from Apache won’t be easy to adapt to Facebook’s unique architecture. From the blog post:

It’s worth noting that we considered Apache YARN as a possible alternative to Corona. However, after investigating the use of YARN on top of our version of HDFS (a strong requirement due to our many petabytes of archived data) we found numerous incompatibilities that would be time-prohibitive and risky to fix. Also, it is unknown when YARN would be ready to work at Facebook-scale workloads.

So, Facebook plods forward, a Hadoop user without equal (save for maybe Yahoo) left building its own tools in isolation. What will be interesting to watch as Hadoop adoption picks up and more companies beging building applications atop it is how many actually utilize the types of tools that companies like Facebook, Twitter and Quantcast have created and open sourced. They might not have commercial backers behind them, but they’re certainly built to work well at scale.

Feature image courtesy of Shutterstock user Johan Swanepoel.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• A near-term outlook for big data
• Dissecting the data: 5 issues for our digital future
• Defining Hadoop: the Players, Technologies and Challenges of 2011

DCS doesn’t sell to just anyone, though. The unit’s banner customer used to be Facebook (although it has since begun building much of its own gear, and claims to be off vendor gear entirely for its newest data center), and others include Microsoft, Salesforce.com and eBay. In fact, Dell provided the bulk of the web servers for eBay’s Project Mercury data center that I profiled in April. Outside the web space, Dell’s DCS customers include large oil & gas companies and research centers.

Those types of customers are important. As cloud computing and large web sites have shifted server-sales dynamics over the years, fewer customers account for an ever-increasing percentage of server sales. Thanks to Facebook’s Open Compute Project and other efforts, they’re increasingly demanding high-density gear that packs as much power as possible into the most-efficient footprint. In fact, the company claimed in 2011 that its DCS division, combined with its off-the-shelf PowerEdge C line of energy-efficient servers, would be the fourth-largest x86 server vendor in the world if it were spun off from Dell.

To give you a sense just how many servers 1 million is, consider that Google runs about a million servers (give or take, if estimates are accurate), while Facebook is estimated to be running about 181,000 servers. Dell has produced an infographic putting the number into context in terms of energy savings, as well.

And you can check this post from Dell’s Barton George for a celebratory video, as well as a photo of the napkin on which the whole DCS idea was hatched.

Feature image courtesy of Shutterstock user Oleksiy Mark.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Migrating media applications to the private cloud: best practices for businesses
• Facebook, Apple Building New Data Centers, But Why?
• A near-term outlook for big data

According to Quantcast VP of Research and Development Jim Kelly, the web-audience measurement specialist began working with Hadoop in 2006 and experienced problems almost from the start. However, while the early problems with HDFS might have been symptoms of its immaturity, the problems soon began centering around the two things Hadoop is supposed to be best at — size and speed. So, in 2008, Quantcast began experimenting with, and actually sponsoring, the Kosmix project.

It turns out that wasn’t a moment too soon. By 2010, after Quantcast began integrating with ad networks, its data flow really began picking up into the tens of terabytes a day range. It turned on QFS as its production Hadoop file system in 2011 and now receives about 40TB a day and processes a whopping 20 petabytes. Kelly said Quantcast has pushed 4 exabytes — or 4 billion gigabytes — through QFS since turning it on.

Faster, yes. Bigger, not so much.

At Quantcast’s scale, the problem with HDFS wasn’t so much its scalability, but the sheer size of the cluster required to handle petabyte-scale data stores. HDFS stores three copies of each piece of data to ensure they’re always available, although it tries to make up for the size issue with data locality (i.e., putting data directly on the computing nodes so it doesn’t have to traverse the network in order to be processed). Kelly thinks those techniques are relics of a bygone era.

“When HDFS [was created], disk drives and networks were tied for being the slowest things in the cluster,” he said.

Enter Reed-Solomon error correction, QFS’s chosen method for assuring reliable access to data that Kelly says actually ends up shrinking the size of Hadoop clusters while improving their performance. (It’s actually the same method used on CDs and DVDs.) Rather than storing three full versions of each file like HDFS, resulting in the need for three times more storage, QFS only needs 1.5x the raw capacity because it stripes data across nine different disk drives. Quantcast believes smaller cluster size, combined with today’s 10-gigabit networks and the ability to read and write data in parallel make QFS significantly faster than HDFS at large scale.

QFS also comes equipped with other features that Quantcast had to implement to make it production-ready. Among them: it is written in C++ and has fixed-footprint memory management; it has access control based on users and groups; and it intelligently detects node failures, as opposed to planned maintenance, and invokes data recovery accordingly.

It’s not for everyone, though

Despite its claimed improvements over HDFS though, Kelly is quick to point out that QFS is probably not the best choice for everyone. It’s really designed for Hadoop users operating at petabyte scale, who have the technical prowess to handle a migration away from HDFS, and for whom data-processing costs are hitting the six-to-seven-figure range monthly once things such as energy bills accounted for.

“If you’re cluster only has 10 disk drives,” Kelly said, “[QFS] will save you $500, which is nice but …”

Likewise, if high availability is very important, the latest version of HDFS might be preferable. “There’s a standby [in QFS]; it’s not quite as hot as theirs,” Kelly said. But availability isn’t super important to Quantcast, he said, it hasn’t had any real problems with QFS going down anyhow. When it does, it actually recovers pretty fast.

As for the rest of the file systems touting themselves as better alternatives for HDFS, Kelly didn’t have much to say. Quantcast’s efforts are focused on mega-scale Hadoop deployments, and he doesn’t see anything better for that use case. Although, he noted, Hadoop vendors probably shouldn’t get too upset over all the competition.

“I think some diversity in the ecosystem is probably not a bad thing,” he said, “and is probably a sign of healthy evolution.”

Feature image courtesy of Shutterstock user Lobke Peers.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• A near-term outlook for big data
• Defining Hadoop: the Players, Technologies and Challenges of 2011
• How search can unlock the power of big data

Don’t get me wrong, Twitter has been nothing if not generous in talking about the software it builds. The company has open sourced numerous data-management tools and other pieces of code. It has occasionally (at least in 2009) been willing to share how it handles, stores, searches, and analyzes billions of data points relating to users and their tweets. But when it comes to the actual infrastructure on which this software runs?

Let’s just say Twitter is less than forthcoming. The explanation of today’s outage:

The cause of today’s outage came from within our data centers. Data centers are designed to be redundant: when one system fails (as everything does at one time or another), a parallel system takes over. What was noteworthy about today’s outage was the coincidental failure of two parallel systems at nearly the same time.

It was minimally more descriptive in explaining the cascading bug that took the site down temporarily last month.

And after publicly detailing the migration into its new data center last March, Twitter has been mysteriously mum about where it’s actually running. In Salt Lake City? In Sacramento? The response from a Twitter spokesperson when asked where it’s data centers are located last June: “I can also confirm that we have multiple sites, but I won’t go into further detail.” Apparently, it now has space in Atlanta, too.

If Twitter wants to remain opaque about its practices, that’s fine — but it shouldn’t expect any slack from upset users or investors. By contrast, we have a pretty good idea where Google’s data centers are and what’s going on inside them, and we know nearly everything about Facebook’s operations. When Amazon Web Services has an outage, it might take days, but provides a detailed post-mortem report explaining what went wrong.

Even if Twitter’s infrastructure team is filled with very smart engineers, there’s certainly benefit to be derived from public discussion about what it might be doing right and wrong. Clearly, something isn’t right; the site is down too often considering how much smaller it is than the aforementioned services. While it’s not disruptive enough to anyone’s business to warrant an AWS-style explanation, we gotta get something better than blaming two hours of downtime on an “infrastructural double-whammy.”

Image courtesy of Shutterstock user Elnur.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Takeaways from the second quarter in cloud and data
• A near-term outlook for big data
• What converged infrastructure means for the future of the data center staff

And it wants to make this transition as easy as possible for customers that don’t want to get their hands dirty working with Hadoop or trying to turn a relational database into something it isn’t.

Drawn to Scale’s product, called Spire, is built atop HBase (a distributed NoSQL database that uses the Hadoop Distributed File System), so it scales with ease. Because it uses SQL, users can write robust queries in a familiar manner. Because it uses a distributed index and knows exactly where to go to find the right data, it’s fast as heck. Because it leverages Chef to configure the database, there’s no need to learn the finer points deploying and managing a Hadoop cluster.

“We’re sort of the only game in town when it comes to cloud-scale databases that run on top of Hadoop,” says Drawn to Scale Founder and CEO Bradford Stephens. There are plenty of companies that have lots of data — petabyes of it, in some cases — and want to find new ways to analyze it, but they don’t necessarily want to learn how to deploy Hadoop or write MapReduce workloads.

As of Tuesday, Drawn to Scale has another carrot to offer prospective customers because Spire now ships with MapR’s M3 distribution pre-integrated as the product’s Hadoop platform. Although it’s the bane of open source devotees such as fellow Hadoop vendors Cloudera and Hortonworks because it uses a proprietary file system in place of HDFS, MapR is gaining quite a following because of the performance benefits its technology brings. In fact, Drawn to Scale is just the latest MapR partner after Amazon Web Services recently tagged it for inclusion in its Elastic MapReduce offering and Google did the same for its Compute Engine cloud.

Bradford Stephens

When you’re building a distributed database, you need a fast distributed file system, he added, and MapR is the fastest Hadoop file system around. Because it’s fully API-compliant with HDFS, though, M3 still works just fine with Spire’s HBase underpinnings. “Despite our small size,” Stephens said, “we are more than willing to pick a horse and really advocate loudly for it.”

Although Drawn to Scale is relatively young — it’s still in private beta and just raised a $925,000 first round in March — large companies are already buying into its message. Already, Stephens told me, it’s engaged in seven paid pilot programs with large credit card companies, telcos and other “traditional” types of enterprises. One of them is already trying to negotiate a deal for a 1,000-node production deployment, he added. The company’s skeleton crew of mostly engineers couldn’t keep up the support workload if it opened the floodgates on all the inbound interest, Stephens said.

That kind of demand isn’t surprising when you consider Hadoop’s promise as the platform for a new generation of data-driven platforms. Whether or not anyone actually wants to use MapReduce — the programming framework that made Hadoop famous — they do want to take advantage of its nearly limitless scale on cheap, commodity hardware. When companies such as Drawn to Scale bake Hadoop into an otherwise useful product as a technological component rather than as the product itself, customers should experience the best of both worlds.

Feature image courtesy of Shutterstock user Jakub Pavlinec.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• A near-term outlook for big data
• With Scalable Data Stores Around, Is NoSQL a Non-Starter?
• Big data 2013: key trends and companies to watch

Granted, EMS’s boxes are nowhere near as powerful as a massive data center chock full of computing gear, but that’s kind of the point. Modular data centers are all the rage right now because they let companies grow capacity as its needed, whether that’s a rack at a time inside an IO Data Centers unit or 1,920 servers at a time inside one of eBay’s specially designed modular units.

But modular doesn’t necessarily mean easy or flexible. At the smallest, a standard shipping container unit needs 120 square feet of floor space and can weigh 100,000 pounds, while one of IO’s units needs 500 square feet.

If you just need a rack full of dense computing power that can go quite literally anywhere you have room, EMS might be your provider. The company’s biggest unit, the 42U and roughly 165-cubic-foot R.A.S.E.R. HD, can handle up to 80 kilowatts at a PUE of 1.1. Its mid-range unit — the R.A.S.E.R. DX, which is the 105-cubic-foot unit AOL uses — can support 12 kilowatts and can fit through a doorway. Its smallest product, the C3-S.P.E.A.R. comes on wheels.

Aside from their size, the micro-modular data centers (as the company calls them) are so versatile because they’re fully contained units complete with state-of-the-art cooling and fire-supression technologies. EMS CEO Bill Stockwell explained their appeal to me like this: “Can you imagine taking a gallon of milk out of your refrigerator, putting it on the counter, and cooling your whole house [from the coolness it puts off]?” That’s how many traditional data centers function, he said, with central cooling that has to handle all the racks within the facility. “We make a refrigerator unit for IT deployment.”

Tony Cole, EMS’s vice president of North American sales, told me many customers put them inside office building in lieu of building out new rooms capable of handling computing gear. Royal Caribbean puts EMS units on its cruise ships. AOL appears content putting them on concrete slabs outside its offices.

Others that have their own data centers will deploy a micro data center when they want to increase density for a certain application but their current infrastructure can’t handle tens of kilowatts per rack. That’s what happened with a missile range in New Mexico that bought a load of high-density gear but then found out it would cost $900,000 to retrofit their data center to support it. Instead, they consolidated 10 racks into two R.A.S.E.R. HD units and actually acquired more computing capacity as a result.

Because it’s liquid-cooled, “there isn’t an IT package on the market today that the R.A.S.E.R. HD couldn’t support from a cooling perspective,” Cole said.

Currently, Cole said, EMS’s micro data centers often act as disaster-recovery sites for mission-critical applications or sometimes get placed at remote offices. But EMS co-founder and Chief Technology Evangelist Simon Rohrich said he sees a shift happening as more companies get interested in owning their own cloud computing infrastructures, but don’t have the budgets or expertise to build and run their own data centers.

In this regard, AOL could turn out to be one heck of an important customer as both a use case and a cheerleader. AOL’s use case of deploying EMS units globally as part of a single cloud platform shows customers how easy it can be to make 10 racks across 5 continents look like a single location with the right software, Rohrich said. And having the company out in public touting 5 times the computing capacity for 10 percent of the cost is sure to turn some heads, too.

Already, he added, customers are starting to realize they can deploy multiple EMS units at a single location to achieve some powerful systems. Cole noted that EMS has deals in place for 13- and 16-unit arrays already, but suspects bigger arrays will come. The company is still in the early stages, he said, “so people are just kind of kicking the tires.”

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• 3 baby steps toward greener data centers
• How a Snapshot of a Green Data Center Can Be Misleading
• How to Make Cloud Computing Greener

Here’s how the company explains the technology in the press release announcing it:

The Extended Memory subsystem dynamically moves frequently accessed data pages into memory on-demand while transparently migrating rarely accessed data pages from DRAM into ioMemory. This allows developers to simplify application design by assuming that entire datasets are in-memory, without the costs associated with DRAM purchase and operation. Application developers are able to further tune performance through software development kit tools that lock selected pages into DRAM, giving access to NAND flash as memory, instead of treating it as an extension of disk storage.

Fusion-io was able to pull this off in part because of an industry trend toward higher-capacity, less-expensive flash storage and in part because Fusion-io doesn’t claim to offer solid-state drives. When I covered the release of the company’s ioDrive2 and ioDrive Duo2 (which plug into a server’s PCI slot and operate as repositories for data needing high performance) in October, the products represented a more than 2x capacity gain over the previous generation and continued to boost IOPS performance, despite the smaller and cheaper NAND technologies on which they’re built.

Fusion-io Founder and CEO David Flynn attributed his company’s ability to continually raise performance on new, consumer-driven NAND designs with its decision to to provide memory controller software rather than trying to act like a hard-disk drive (as other solid-state drive providers try to do) and use a microcontroller. That same decision, the company says, also helped make Extended Memory possible because it didn’t have to try and make disk-storage protocols function at the memory tier.

In-memory databases and analytic systems (such as SAP’s high-performance HANA appliance) are nothing new, but they’re limited in size to the system’s memory footprint and can be rather expensive. All-flash storage arrays are also becoming much more popular and getting less expensive, but they still exist as a separate storage tier and are often aimed at legacy enterprise applications. If what Fusion-io and Princeton have created works as promised, it could create a middle ground approach that resonates with customers — such as Fusion-io mega-user Facebook — that need a workable combination of massive scalability and high performance.

Image courtesy of Shutterstock user Mark Schwettmann.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Infrastructure Q1: Cloud and big data woo enterprises
• A near-term outlook for big data
• Infrastructure Q4: Big data gets bigger and SaaS startups shine

The Nibiru project, he explains, is a set of “incredibly game-changing” goals for transforming the way AOL’s services division carries out the work of managing the company’s infrastructure, and the newly materialized mini data centers we’re high on the list:

Our primary “Nibiru” goal was to develop and deliver a data center environment without the need of a physical building. The environment needed to require as minimal amount of physical “touch” as possible and allow us the ultimate flexibility in terms of how we delivered capacity for our products and services. We called this effort the Micro Data Center. If you think about the amount of things that need to change to evolve to this type of strategy it’s a bit mind-boggling.

Among those changes, Manos writes, are capabilities such as being able to deploy infrastructure wherever it’s needed regardless of temperature and humidity, and the “ability to fit into the power envelope of a normal office building.” The units will also be part of AOL’s automated cloud computing infrastructure, which means they’re managed as part of a greater pool of resources without the need for dedicated staff.

But these mini data centers aren’t just about a cool operations project; they could end up paying big dividends for AOL’s business lines, as well. Not only can AOL save money by taking up less space and power within a traditional colocation facility — if it decides to place a Nibiru box in such a facility at all — but the box gives AOL the flexibility to move into new geographies with ease. Among the benefits Manos points to are:

• It allows us an incredibly flexible platform for driving and addressing privacy laws, regulatory oversight, and other such concerns allowing us to respond rapidly.
• Gives us the ability to drive Edge Computing delivery to potentially bypass CDNs for certain content.
• Gives us the capability to drive ‘Community-in-a-box’ whereby we can quickly launch new products in markets, quickly expand existing footprints like Patch in a low cost, but still hyper-local platform, allow the Huffington Post a platform to rapidly partner and enter new markets with minimal cost turn ups.

Containers atop eBay’s Project Mercury data center in Phoenix.

AOL’s new boxes might be about the smallest data-center-in-a-box units around, but they’re actually part of a great move toward modular data centers across companies of all types. As I reported in April, eBay is buying custom-built containers full of thousands of servers that it can drop into (or on top of) its data centers as capacity dictates. IO Data Centers has built an entire business around modular data centers that can sit just about anywhere.

Update: AOL’s micro data centers were built by Elliptical Mobile Solutions, according to a spokesperson for that company. It has also built units, which it calls micr0-modular data centers, for NATO, the U.S. government and the Canadian Department of Defense.

All these efforts share the same goal of letting companies grow their server count when and where needed rather than trying to predict necessary capacity and relevant geographical locations years in advance.

Related research and analysis from GigaOM Pro:
Subscriber content. Sign up for a free trial.

• Big Data, ARM and Legal Troubles Transformed Infrastructure in Q4
• Social third-quarter 2012: analysis and outlook
• The economics of clean-data-center innovation