HPC’s history is in massive — and massively expensive — monolithic servers housed in the bowels of immense machine rooms deep inside the federal government and corporate research facilities.

And that still is one face of HPC, but that’s not where the growth is. There is a lot of new money pouring into HPC, and the promise of much more to come, in smaller clusters of 32 to 128 nodes. In this article I’ll look at what this new market is about, and at how some of the vendors are adjusting to meet its demands.

Enter the workgroup segment

High performance computing is the fastest growing segment of the IT industry. The market for HPC servers grew from about $5 billion to just over $10 billion between 2001 and 2006 according to IDC, and in 2006 26 percent of all processors sold into the server market were were installed in HPC servers. This growth is happening even as sales in the very high end of the market (especially systems priced $5M and up) have actually declined in recent years, falling by as much as 29 percent in 2006 alone according to IDC.

The real growth in HPC right now is in the sale of smaller systems.

According to Steve Conway, a research vice president for HPC at IDC, “The fastest growing piece of the HPC market is for systems priced under $50,000. A rash of new products in this price range have been coming to market, and the watchword is ‘ease of everything,’ from purchasing to installation, deployment, management, and expansion.”

For the sake of convenience I’ll call these “workgroup” HPC systems, regardless of whether they are used in a small business, an academic department, or in an actual workgroup. We ended 2007 with all of the major HPC vendors having staked a position relative to this new workgroup market.

At one extreme is Cray, exclusively committed to the high end of the scientific HPC market. At the other end is Linux Networx, focused on systems in a $50k to $500k sweet spot. Then there are those straddling both markets. Companies like HP, IBM, and SGI are adding layers to their existing services and strategies in an attempt to hold on to their heritage on the “super” end of supercomputing, while at the same time diversifying to try and capture some of that new revenue.

With a low end like this…

With the high end actually shrinking, you might be tempted to ask why a company would bother with it at all?

The problem is that the path to success in the HPC workgroup market is not clear. Much of the potential of this market is buried under the reality that its customers don’t identify with HPC technologies as part of their solution set. At the same time HPC companies have to figure out how to position their technologies in terms relevant to this new market. It will take time and experience in each of the many new markets that workgroup HPC is trying to penetrate to create solutions that connect with customers.

Finally, the software isn’t there yet to really make mass market HPC possible. Accelerators may offer an interim solution if the ISVs invest in having their applications take advantage of them on the way to deeper parallelism; IBM, HP and SGI in particular are making significant efforts to integrate accelerators of various sorts into their offerings. But getting to a critical mass of robust parallel commodity software will take time.

Getting to Volume

In order for this market to be profitable, companies have to do a much larger volume of business in smaller systems. Building volume will obviously be easier for those companies that already know how to market on a mass scale, or already have broad connections in the enterprise world. Among the clear favorites here are IBM, HP and Dell, by virtue of their heritage as truly mass market companies.

Dell’s DNA is in getting large volumes of product out the door to a variety of customers, ranging from corporations to individuals. IBM and HP have a strong history of direct sales in the enterprise, with IBM present at the birth of much of what is now called enterprise computing. As the market evolves we may see that these two companies have a slight edge with those HPC customers having an “enterprise IT” mindset, with Dell leading in those purchases that are more individually-oriented on the strength of its direct-to-consumer marketing heritage.

Getting heard

There’s having access to a large volume of customers, and then there’s having something to say that makes them open their wallets. Watching the vendors try to understand these new customers and then try to position their technology as relevant to them will be a fascinating exercise to watch as it unfolds during 2008.

Some of the vendors are focusing on domain-integrated solutions. Linux Networx, for example, offers pre-configured solutions in industry verticals (like automotive, oil and gas, and so on) that come tuned to run applications relevant in the vertical. Dell is working with ISVs to construct system configurations that are specifically designed to support customers solving real problems. When these configurations are marketed to customers they are described in terms relevant to the problems customers are trying to solve, rather than in abstract processor counts and memory sizes.

HP is hoping to get to customers in this new market through partners that have existing connections to customers already using specialized analysis tools on workstations (like Fluent or Accelrys Materials Studio). Software resellers will be able to offer configurations with their software pre-installed, and tuned to support their applications.

As Dan Reed, director of Scalable and Multicore Computing Strategy at Microsoft points out, this kind of integration will be key to winning customers who don’t want a supercomputer per se, they just want to get work done. Says Reed, “Vertical integration is the name of the game in the smaller market for people who want to own/lease their own iron.” As Dr. Reed points out, it will also be key to offer services to these customers that help them with the transition into HPC, “Finding qualified people who understand technology, applications and business (all three) is a challenge that limits broader adoption, particularly for small and medium sized companies.”

Getting to “yes”

Once you’ve got a hundred thousand prospects, and you have something to say that they want to hear, you’ve got to offer them something they can use. Yes, a hundred gross of rocket launchers probably would enable the stray dogs of New York City to stage a coupe, but they don’t have opposable thumbs. No sale.

New workgroup HPC customers are going to have to have their pain buying, deploying and running their new hardware salved before this market can really take off.

Both Dell and HP are betting strongly that the path to profits on workgroup-sized clusters is enabling customers to help themselves. Though Dell probably has the edge here, both companies have a legacy of mass sales through online configuration and ordering tools, and both plan to bring this expertise to the HPC market with tools to help customers build clusters to solve their own problems. HP’s Cluster Platform Workgroup systems even come in starter configurations that plug into a standard 110V wall socket.

Linux Networx is taking a “mass customization” approach, offering customers its Validated Performance Engineering (VPE) process that provides for a structured interaction between Linux Networx solution professionals and customers to come up with a solution that’s right for them, tuned for their needs, and ready to run very quickly.

All of the vendors are funding programs to layer system management tools into their offerings, and some are adding reliability and recovery features. As an example here, SGI is developing its Industrial Strength Linux Environment (ISLE). ISLE is meant to provide a solid operating foundation and then be combined with advanced graphical front-ends and management tools that make running the clusters more intuitive.

IBM is pushing some of its virtualization technologies up to its HPC machines in a effort to bring popular enterprise features to that product line. And a notable industry effort aimed at giving customers the confidence to buy is the Intel-sponsored Cluster Ready program. ICR enables customers to know that a given software stack will run with a given hardware configuration.

But infrastructure and usability issues are still likely to loom large for these potential new customers, accustomed to ordering a workstation and plugging it in. And, even with the reliability improvements in progress for modern HPC systems, there are still far too many “what the heck just happened” moments in clusters of all sizes.

With the benefit of a decade or so behind them, the vendors will probably agree that a slow start to the HPC workgroup market is a good thing. HPC still needs time to figure out how to offer a “plug and compute” solution that works at least as well — from purchase to data storage — as the modern PC does today, and preferably better.

—–

John West is a freelance technology writer, a frequent contributor to HPCwire, and the publisher of insideHPC.com. He’s been involved in supercomputing for the past 15 years.