The cloud computing meme is permeating practically all areas of computing these days, including HPC. Will the cloud replace the grid as the new paradigm for delivering high performance computing? To be fair, it’s not that grid computing never delivered; it just never reached escape velocity for the HPC market.

With the emergence of commercial solutions, like Amazon EC2, Google App Engine, Sun’s Network.com, and IBM’s Deep Computing on Demand, cloud-based utility computing is starting to look a lot more mainstream. Not wanting to be left out of “The Next Big Thing,” HP, Dell, Microsoft, Yahoo, and others are all jumping onto the cloud-wagon — if you’ll pardon the mixed metaphor.

So will the cloud swallow HPC just as commodity computing machines did ten years ago? Probably, but it’s too soon too tell how fast that will occur. The technical challenges — networking bandwidth and latency, compute performance, and software standards — are still an issue for HPC but are quickly being solved for enterprise computing. And the business model for large scale utility computing is still being worked out.

Cloud computing maven Nick Carr has some ideas about that. In a recent post on his blog, Carr says organizations like Amazon or Google have the edge right now. His hypothesis is that companies that have already built an ultra-scale compute infrastructure for their low-margin retail or advertising business can leverage that investment into a higher-margin utility computing business. “For Amazon, running a cloud computing service is core to its business in a way that it isn’t for, say, IBM, Sun, or HP,” explains Carr.

How does that translate to HPC? Ultra-scale platforms like Google and Amazon are not really set up for HPC apps at this point. The only large scale supercomputing infrastructures are owned by governments, TeraGrid in the U.S and DEISA in Europe being the best examples. They have no way to deliver all that compute and storage capacity as a commercial utility solution and, being research-oriented, have no mandate to do so.

Outside of the cloud model, individual HPC systems can be farmed out. The DOE, for example, shares its high-end supercomputers with industry (and academia) via its highly regarded INCITE program, but not as a commercial service. The lucky few companies that win the INCITE lottery get to use the cutting-edge supers for high-end industrial research, but not for day-to-day computing. While government-industry HPC partnerships have become rather common, the model isn’t geared for production work.

In the commercial arena, IBM’s Deep Computing Capacity On-Demand (DCCoD) rents out Blue Gene cycles as well as capacity on less exotic platforms from its DCCoD centers. And Sun’s Network.com lets customers buy compute time by the CPU-hour on a modest-sized x86 cluster, while also offering access to a handful of HPC application suites. The long-term viability of this model is still a question.

More in the Nick Carr model of doubling up on in-house computing resources, the Computational Research Laboratories (CRL) in Pune, India, is offering up Eka, its new 117.8 teraflop supercomputer for commercial use. As the number four system on the current TOP500 list, Eka is the only privately owned supercomputer in the top 10. CRL itself is owned and operated by the Tata Group, and according to a Financial Express report, the machine will be used by the group’s Tata Motors (automotive) and Tata Elxsi (product design) subsidiaries. By also offering the $30 million machine as a supercomputing service platform, Tata intends to get the most return on its investment.

An IEEE Spectrum article published today talked about Tata’s CRL HPC business strategy:

True to India’s software and services tech culture, rather than try to outdo Cray, IBM, Hewlett-Packard, or Silicon Graphics at designing and selling supercomputers, CRL will provide end-to-end supercomputing services—renting computer time, adapting and fine-tuning applications, and offering analytical services. Today Eka is testing more than 15 applications for customers, and the company is in talks with several clients from the automobile, aerospace, financial, oil and gas exploration, and life sciences sectors, including aerospace giants Boeing, Embraer-Empresa Brasileira de Aeronautica, and Airbus.

Even if this arrangement proves to be workable for Tata, the model would be hard to reproduce. How many commercial organizations can afford to buy supercomputing resources at a scale that would serve a reasonable number of HPC renters and offer the kind of software support that CRL is intending to provide? Besides Tata, I can’t think of anyone else.

Which brings us back to cloud computing. The way I envision HPC moving over to the cloud in mass is when the aggregate performance of the systems is so great that the lack of efficiency won’t really matter. These systems will be able to “waste” compute, storage and network resources because economies of scale will render monolithic machines way too expensive. We’ll know that has happened when NCAR is running their climate models on Google EarthSim and Boeing is designing airplanes on Amazon WindTunnel.