There is much conversation these days about parallel programming, as the broader development community takes an interest in getting at the power of the multiple cores that seem destined to park themselves under consumer desks and on traveller laps around the world. You can find both declarations of doom and promises of hope in those conversations. Parallel computing is variously described as too hopelessly difficult for any but a few to ever practice it effectively, or about to be revolutionized by tools that will make it “easy” for anyone to program in parallel.
One thing that no one is seriously talking about being “easy” anytime soon, however, is selecting, buying and installing compute clusters. At the middle and low end of this market, deploying clusters is difficult. Users lack experience with acquiring moderately-sized equipment, installing and configuring it, and keeping it running once the installation is complete. These challenges are at odds with market forces — identified by research firms like IDC, Tabor Research, and others — that portend tremendous growth for high performance computing at a smaller scale. And the distance between a market’s desire and what’s commercially available creates an opportunity for new products.
In the United States the HPC market is fairly diverse on the provider side. There are major HPC vendors with deep infrastructure and expertise to support what are essentially national-scale supercomputers. We also have a variety of smaller participants specializing in clusters for specific groups of users — enterprise users hosting large multimedia databases or Web servers, for example — that can be deployed with little customer expertise because of the added value in software and system engineering offered by these niche vendors. The market is rounded out by the middlemen, systems integrators and the like, who in the US can step in to bridge the gaps left in this market. They work with consumers whose needs aren’t directly addressed by one of the niche providers, but who are still too small to get the focused attention of the major vendors.
In Europe, and the United Kingdom in particular, the market has a different shape. As I was talking to Julian Fielden, the managing director of HPC provider OCF in the UK, he made the observation that systems integrators and what you might think of as “third party” providers of small-scale HPC equipment in the United States play a relatively smaller role than in the US market because of the lack of these niche providers in the UK. Smaller companies like OCF have then grown up in this environment to provide, as Fielden puts it, “the close-to-the-customer experience and high levels of expertise” that the major vendors simply don’t have the resources to supply to the many hundreds, or thousands, of smaller purchasers.
Based in Sheffield, UK, OCF started in 2002 as an independent consultancy building custom HPC solutions for their customers. “This model worked for a while,” says Fielden, who founded the company, “but then IBM launched its tier 1 Opteron box, and it was winning everything.”
As OFC reassessed its business model and market opportunities, it still recognized the need to provide higher touch services to smaller customers than the major hardware vendors were able to provide, but it also recognized that competing head-to-head on hardware with a company like IBM was a losing proposition. So rather than compete, it partnered, using the experience and expertise it had built up as as a totally independent provider to help customers through the difficult process of building, maintaining and using HPC solutions built on IBM server and storage technology.
OCF isn’t the only integrator that IBM is working with, and IBM isn’t the only company using this model in the UK and Europe. Another company taking a similar approach to meeting the needs of Europe’s HPC customers is ClusterVision. With offices located throughout Europe and the UK, ClusterVision also partners with major HPC providers, like IBM and Dell, to provide the same kinds of services that OCF provides small HPC customers. I spoke to Matthijs van Leeuwen, director of sales at ClusterVision, in my quest to better understand this part of the European HPC market dynamic.
The first thing I wanted to know from both men was what are they doing about the elephant? How do they manage to get along with the very large companies whose products they are selling when those large companies also do their own direct business? ClusterVision’s van Leeuwen says that the relationships are managed first of all with “good communications and honesty.” The best of intentions can go awry, however, and ClusterVision’s sales and presales teams work closely with their partners at all levels of a deal, and follow specific rules of engagement (which were not for public consumption) to help avoid situations where the companies would stomp all over each other’s toes.
Fielden’s OCF takes the same sort of approach, maintaining close relationships with IBM and staying in contact through their partner channels throughout the lifecycle of a deal. He describes the relationship these vendors have with their partners as “coopetition,” with both companies recognizing the area of the market they serve best and acting together to preserve their overall best interest.
OCF builds an engagement team with its partners in which OCF internal staff works directly with the customers to make sure that everyone involved works together to build the solution. Although the particulars vary among engagements, OCF is often the “single bellybutton” that customers push to get issues resolved. This makes things easy for customers, who only have to build and exercise one relationship. At the same time, the close relationships that both OCF and ClusterVision have with their partners ensure that the best expertise the partner company has to offer can be made available to the customer. The clear benefit to the larger vendors is that their gear gets sold and supported without them having to build out and staff a business scaled for thousands of small customers.
Both companies see the HPC market continuing to have dramatic growth over the near term. “There is no point in looking 5-10 years ahead, because there is just too much change in the market,” Fielden says. “But over the next 3 years the market should keep growing at a rate of 30-40 percent.” For Fielden this opportunity has a challenge baked in. The market growth will continue to expand OCF’s top line, but the company has to be managed carefully to make sure that growth falls through to the bottom line.
ClusterVision’s van Leeuwen also sees opportunity, “We see the HPC market in Europe growing, in all sectors. ClusterVision aims to grow with the market and beyond… and will grow as fast as we can without compromising on quality.”
Both companies have a diverse offering that allows them to add value for their customers beyond the system they buy. OCF can work with customers at the very beginning of the acquisition process as a sort of trusted advisor, helping to understand and articulate an organization’s needs and identify the right opportunities for HPC. They then design systems for storage, data visualization, and computation to meet those needs. After the sale, OCF can provide support, system administration and data management to meet specific customers’ requirements, and can work with customers to tune and port application codes and provide high-level computational science expertise. Fielden describes OCF’s staff as being “as expert as IBM in most of their technologies, but more immediate and accessible” to customers.
ClusterVision builds, markets and sells its own branded clusters created from commodity (i.e., “whitebox”) components, as well as reselling Dell, IBM and Supermicro gear. On both sides of the sale, the company provides much the same management and computational science expertise that OCF does. ClusterVision does, however, operate at a larger scale than OCF and has many more employees since it provides services across the whole of Europe. But where ClusterVision has taken a slightly different business direction is with its ClusterVisionOS, an HPC software stack that productizes the company’s HPC knowledge into a ready-made package that simplifies and standardizes the creation of a working cluster.
According to van Leeuwen, the major software components needed for a fully functional cluster are an OS (usually Linux), along with configuration tweaks to support the cluster environment; 3rd party open source (MPI and so on); 3rd party commercial software (compilers, etc.); and management software. “The ClusterVisionOS has all of the above fully integrated,” explains van Leeuwen. “For us it is very important to be able to install a complete cluster almost at the push of a button and have consistency in the cluster installation process and the finished turnkey product. Our experience is that this is much better achieved with an integrated solution.” ClusterVision also offer ClusterVisionOS back to vendors who include it in their offering. Dell, for example, offers ClusterVisionOS to some of its customers.
Because they offer their own branded solution, I asked van Leeuwen how they go about deciding when it is best to offer their own products to customers and when it is best to offer a solution based on partner gear. “This results from a sometimes complex analysis of various factors,” says van Leeuwen. According to him, sometimes customers prefer a specific vendor, while at other times a particular customer is important to a particular partner who will then offer a deeper discount to get their gear into the solution.
While they take different approaches in the particulars of their businesses, it is clear that both OCF and ClusterVision are integral to creating a market for HPC in the UK and Europe that reaches from the top to the bottom of the customer space. As the HPC market continues to rapidly change shape over the next five years or so, it will be fascinating to watch where the line of coopetition between the big players and their supporting integrators shifts.