RAM | Reseller Advocate Magazine: Issue 55

By Chris Angelini

THE IDEA OF working on a supercomputer may seem somewhat daunting. After all, real supercomputers are ranked in order of their processing horsepower and compared to 500 other titans around the world. You may find servicing one to be more confusing than finding the space shuttle's gas cap. But that's probably because most folks still think of room-sized behemoths when they envision the giants of computing.

Modern supercomputers often don't fit that stereotype. They've given up proprietary vector processors, commonly used in the '70s, in favor of less expensive commodity components running in massively parallel configurations. High-speed interconnects such as InfiniBand conjoin each piece of the supercomputer, facilitating high-bandwidth, low-latency data access for demanding loads.

You can see where this is going—today's supercomputers aren't the impossibly complicated rooms of whirring tapes and flashing multi-colored lights they once were. Instead, they're built using many of the same parts that go into the white box servers you already sell, running in cooperative configurations. That's where we get the concept of clustering. And it might surprise you to learn that resellers are starting to explore the potential opportunities in clustering.

Defining a Need

Most of your SMB customers are probably using single- or dual-processor servers for their day-to-day computing needs and getting along just fine. Few of them need more than a server or two. And there's a good chance you can't think of a single business looking to buy a cluster of servers. But you'll find an unlimited need for processing horsepower in a few specific markets.

"The engineering and science segments are constantly looking to improve processing capabilities," says Stephanie Davis, the senior product manager of high-performance computing at Microsoft. "Specifically, the automotive and aerospace industries use a lot of parallelized CAE (computer aided engineering) software, as do the oil and gas companies—the geosciences folks. We also see potential for high-performance computing in financial services risk modeling software and business applications."

Self-Contained Clustering
Tyan's Typhoon houses four separate nodes, each with up to two processor sockets and 64GB of memory. Despite overwhelming muscle and a very sleek looking chassis, the Typhoon is surprisingly affordable and easy to configure.

Ching-Ching Ganley, Tyan's high-performance computing product manager adds, "Any application optimized for parallelized operation will benefit from running on a cluster. We're also focusing on the scientific community but believe education is another great place to sell clusters. Small clusters are easily able to replace larger, rackable 1U arrays in schools." Microsoft's Davis clarifies that you won't be selling clustered servers to perform database failover duties or to serve as an Exchange Server store. You roll out a cluster when the customer requires true high-performance computing.

The Hardware is There

As you can imagine, there are plenty of professionals who'd love to spend time with a supercomputer, crunching through equations, rendering complex scenes, or debugging advanced code. Thus, time on the world's fastest computers has to be scheduled. And if a customer's software doesn't compile successfully because of a bug, for example, it's back to the drawing board.

There aren't many organizations with the resources to procure that level of horsepower within their own walls, making the clock a painful equalizer. But if you were able to offer supercomputer-ish levels of computational capacity to customers with a need for speed, perhaps you could soften their reliance on someone else's hardware. Of course, then you're back to the issue of working with highly integrated hardware and rolling out special software. Fortunately, both are now easily accessible to the VAR community.

Worry-Free Deployment
The Typhoon isn't just about fast hardware. It’s also about a complete cooling solution, the ability to plug into a standard wall socket, and not needing a separate chassis.

Tyan is using clustering as a means of giving its customers more computing horsepower than anything previously available, calling its Typhoon offering a personal supercomputer. The idea is that, backed by enough muscle, those same professionals can debug code, render graphics, or crunch through equations at their own desk. Then, when it's time to get the job done, they're able to schedule time on the bigger machine without worrying about a bug stopping the show.

The Typhoon is available in two configurations: The B2881 supports AMD's Opteron processor while its B5160 centers on Intel's dual-core Pentium D. Both models are clusters in that they aggregate the power of four separate 1U servers (called nodes) in a single enclosure.

Each node is plenty powerful in its own right. For instance, the B2881 accommodates dual-core Opteron HE low-power processors. With all four boards sporting two sockets, you can effectively sell a 16-core cluster. You'll find eight 184-pin memory slots per motherboard, giving each node up to 64GB of memory capacity. Moreover, a Broadcom dual-port Gigabit Ethernet controller helps ensure fast node-to-node communication.

Best of all, Tyan designed its Typhoon series intelligently. The nodes are built to fit within a 350W power envelope. Why the artificially low ceiling? When you multiply 350W out by four—the total number of nodes—you get 1,400W. A standard three-prong, 15A electrical socket is generally able to deliver 1,500W of power, so despite housing four dual-socket servers and up to 16 cores worth of computational firepower, the Typhoon doesn't need any special infrastructure work in order to get it running.

The four nodes in each Typhoon cluster are, by default, configured to communicate across Gigabit Ethernet, a cost-effective solution that enables plenty of bandwidth for applications less sensitive to latency delays. However, understanding that some of your customers might be willing to spend a little extra to cut back on wait time, each node also comes with a PCI-X expansion slot you can use to add InfiniBand technology.

Both the AMD Opteron and Intel Pentium D Typhoons will be available in the June/July timeframe. And while speculators have insisted that such a powerful, well-integrated product might run your customer in excess of $30,000, Tyan's John Nguyen says the barebones model, ready for a VAR's touch, should cost roughly $6,000 while fully-configured models duck in under $10,000. Price out four separate servers of comparable pedigree and make the comparison. Then remind your customer that all power, cooling, and storage worries are fielded by that one solution.

Cluster Access, MS Style
A cluster built using Microsoft's WCCS 2003 will sit on an existing Active Directory domain, visible to any connected workstation. Users will only interface with the head node, though, as the compute nodes are tasked solely with processing performance.

The Software is, Too

A complicated hardware environment isn't the only reason VARs would want to avoid selling clusters. Software is also an integral variable in the equation and, up until now, most of your options have centered on a Linux distribution. Unfortunately, none of them are particularly user-friendly, and a lack of unification means it's the reseller's job to pick the right pieces. Certain builds are better suited to particular tasks. And as Microsoft's Stephanie Davis points out, rolling out and configuring a Linux cluster is complicated by different operating system kernels, schedulers, and MPIs (message passing interfaces), never mind the issue of management.

As good timing, or perhaps strategic planning, would have it, Microsoft is putting the finishing touches on its own clustering product called Windows Computer Cluster Server 2003, which incorporates each mandatory slice of the clustering pie in a package supported by one company. It's not Microsoft's first cluster-oriented offering, but it is the first solution specifically written to pull clustering within the reseller's reach.

"VARs will immediately be turned on to the familiar interface of WCCS 2003," says Microsoft's Stephanie Davis. "Moreover, their customers, who are likely accustomed to Windows XP, will have little trouble navigating around WCCS 2003."

Installation is straightforward enough. Windows Compute Cluster Server 2003 is first rolled out on the head node, adding it to an existing Active Directory domain (that's where you get your security settings). You can then use RIS (Remote Installation Services) to deploy the compute nodes automatically. Microsoft's preliminary documentation is explicit in that WCCS 2003 was developed to keep things specifically easy with an appliance-like setup yet flexible enough to guarantee security through authorization mechanisms and the ability to execute scripts.

Once running, resellers with experience rolling out Windows Server 2003 should recognize the operating system's Start Page, used to monitor compute nodes and job information, the To Do List, which helps configure the cluster's various subsystems, and a performance monitor to keep tabs on processor statistics. A couple of new pages—Node Management and Remote Desktop Sessions—displays information on node tasks and facilitates remote connections to the compute nodes since they usually don't have their own displays.

Interestingly, WCCS 2003 requires a 64-bit platform in order to run. Microsoft's Stephanie Davis says that's because the corresponding hardware is no longer futureware; it's here today. Besides, you'll probably soon exceed the 4GB memory ceiling of 32-bit architectures.

Get Ready, Here it Comes

According to Microsoft's Stephanie Davis, WCCS 2003 is still in beta. However, it'll be released to manufacturing in the June timeframe with a price tag in line with the other Windows Server solutions. By then you should expect to have access to the hardware and software needed for incredibly powerful—yet easily manageable—clusters. And although the market for personal supercomputers might be limited, VARs who get in early will face little competition as everyone else experiments with the utility of that much power.