As IT budgets have gotten squeezed, more customers are looking at cloud computing as a way to avoid up-front capital costs, while getting access to as many CPU cycles as they need. In response, all the big IT firms are scrambling to develop a cloud computing product and services strategy, and IBM is no exception.

IBM has actually enjoyed a bit of head start at this. The company’s Deep Computing on Demand offering was launched back in June 2003, when everyone thought clouds were just fluffy white things in the sky. The original offering allowed HPC customers to rent remote access to supercomputer-type systems maintained by IBM. The initial infrastructure consisted of a Linux cluster of xSeries servers housed at a facility at the company’s Poughkeepsie, New York plant.

One of the first users of the service was GX Technology Corporation, a company that does seismic data imaging for the oil & gas industry. Besides dodging the expense of a cluster build-out, one of the big advantages of the on demand service was that the image processing turnaround was much quicker, since IBM could provision up to a thousand servers at a time, depending upon job size.

In general, the original Deep Computing on Demand service was designed for HPC applications across government, academia and industry. Over the next six years, IBM’s on demand offering evolved into a more general-purpose service, broadening its scope beyond traditional HPC, but keeping its computationally-intensive theme. Today it’s just called Computing on Demand and is run more like a cloud with the ability to create virtual images within individual servers.

David Gelardi, IBM’s vice president of Systems and Technology Group for Worldwide Client Centers, sees their current on demand offering as one of the ways in which a client can take advantage of cloud computing today. “In some sense you could think of Computing on Demand as almost a dress rehearsal for cloud,” he says. “We just didn’t know it.”

Currently, there are six IBM on demand centers strung across the US, Europe and Asia. In most cases customer data is stored locally, so bandwidth and latency dictates that the remote servers not be too remote. Because of that, the centers have tended to migrate to “centers of opportunity.” For example, when the oil & gas industry was booming, IBM maintained a center in Houston. As financial services got hot, they expanded into London and New York. Their newest center is in Japan.

Today the two most active sectors of IBM’s on demand service are the financial services industry and industrial design/automation. In the financial space, the applications that support risk compliance plus the creation and management of new types of financial instruments are the two big drivers right now. In the design space, one of the biggest clients is IBM itself, which periodically rents cycles to do large verification runs on its in-house integrated circuit designs.

The six centers currently house a total of 13,000 processors and 54 terabytes of storage. Customers are offered a choice of hardware: IBM Power CPUs (System p servers), or x86 CPUs (BladeCenter and System x servers) using either Intel Xeon and AMD Opteron processors. On the x86 side, both Linux and Microsoft Windows is supported, while the System p users get their choice of Linux or AIX. IBM-built management software, like the xCAT (Extreme Cluster Administration Toolkit), is layered on top for extra functionality.

At one time, IBM offered remote access to Blue Gene technology, but that’s no longer the case. Gelardi says they couldn’t find a broad enough market for the type of specialized technology and support inherent in a rent-a-Blue-Gene offering. The same goes for the Cell processor. He does, however, see the possibility of incorporating IBM mainframes into the on demand model since these represent fairly dear cycles when customers are getting ready to deploy a mainframe application into production.

As far as pricing goes, there are a number of factors that determine cost, including service commitment, technology requirements, and number of compute cycles. It’s actually quite similar to renting other types of infrastructure, like hotel rooms or cars. If you rent for a day, you get one price, for a week, you get a better deal, and so on. Similarly, you get charged a premium if you rent the compute equivalent of a Ferrari versus a Ford. Customer flexibility related to the Service Level Agreement (SLA) is also a consideration. For example, if a customer needs a 24/7 uptime, that’s going to drive the price up since spare servers have to be set aside to account for the inevitable hardware failures.

Gelardi noted that the $1 per CPU-hour for Sun Microsystems’ now defunct Network.com utility computing service might have sounded good, but was an unworkable business model. At some level, he probably wishes the Sun model would have succeeded since it would have kept prices up for all the players. “If I could get a dollar per CPU-hour, I could pave the roads with gold bullion,” he jokes.

Although the IBM compute service has grown beyond its rent-a-supercomputer roots, it still represents a fairly typical compute utility service. The plan, though, is to evolve into a more complex model, where customers will be offered four different types of cloud infrastructure: compute clouds, development clouds, test clouds, and storage clouds. The current offering will naturally evolve into the compute cloud, but IBM’s intention is to develop purpose-built infrastructure aimed at the other three functions.

IBM is already working on a proof-of-concept project with a large financial institution that is looking to give up to 10,000 programmers the ability to independently develop a database plus application service engine in the cloud. The idea for the developer is to be able to attach their workstation to a virtual machine that represents a much larger system. They will also have the ability to do a refresh, which resets the virtual machine back to its initial state.

Although IBM doesn’t supply hard numbers about the size of its computing on demand business, Gelardi says they have hundreds of clients that are currently active or have been active through the course of the program. When it started out in 2003, he says the service was generating revenue on the order of millions of dollars per year. At this point, he says, that has risen to tens of millions of dollars. “As we start to bring in the other types of clouds — the test clouds, development clouds, storage clouds — we’ll blow through the next level very quickly.”