In this week’s issue, supercomputing veteran and Microsoft newbie Dan Reed writes eloquently about two topics close to our HPC-loving hearts: cyberinfrastructure funding for big science and the promise of HPC outsourcing.
First, Reed laments the failure of the recently passed FY2008 Omnibus Appropriations Bill to allocate enough money to fulfill the ambitions of the last year’s COMPETES Act. The 3,500-page bill has been widely criticized as an unsavory combination of pork-laden earmarks and Congressional capitulation to the administration. In some cases, the funding for big science doesn’t even keep pace with inflation, effectively killing COMPETES’ goal of doubling money for the NSF, NIST and DOE Office of Science over the next seven years.
In this new economic reality, Reed also wonders if the research community should start looking for ways to outsource a much greater share of their computing operations to large-scale commercial IT entities. Writes Reed:
“I view this as the research computing equivalent of the fabless semiconductor firm, which focuses on design innovation and outsources chip fabrication to silicon foundries. This lets each group — the designers and the foundry operators — do what they do best and at the appropriate scale. Most of us operate HPC facilities out of necessity, not out of desire. They are, after all, the enablers of discovery, not the goal.”
That notion is also being reflected in the commercial HPC realm, where some users are wondering if utility computing, embodied by the new moniker “cloud computing,” will deliver on the promises grid computing made over a decade ago. John West recently penned a few thoughts about outsourcing HPC, noting that the decreasing cost of FLOPS and the increasing costs of system deployment and IT operations are powerful incentives to move toward a utility model.
Developments along these lines seem to be moving rather quickly in the broader computing community. In 2007, computing giants like IBM, Google, Yahoo, Amazon, and even Microsoft, started to make earnest plans for this brave new world. Web-delivered services and web-based applications are all the rage and poised to become the basis of the next general-purpose computing platform. Once the cloud starts swallowing business and consumer desktop applications, there will be no stopping this trend.
Nicholas Carr’s new book, “The Big Switch: Rewiring the World, from Edison to Google” describes a world where utility computing is quickly reshaping the IT landscape. Carr, who once called the PC “an anachronism waiting to happen,” believes companies like Google, Salesforce.com and Amazon are in the best position to take advantage of this paradigm shift, but that for legacy companies like Dell and Microsoft, it’s going to be a rougher ride.
Carr’s thrust is that computing has become a commodity, just as electricity did in the early 20th century. At that time, a few entrepreneurs realized that electricity could be generated much more cheaply from a large-scale central facility, rather than within individual businesses (users). Electric utility companies were formed and businesses ditched their own power generators and switched over to the more efficient grid.
Today, the cloud represents the nascent computing grid of the 21st century. Businesses that have maintained their own IT departments would willingly outsource their computing if it could be done less expensively elsewhere. While today’s CIOs might be somewhat resistant to turning their company’s datacenter into an exercise room, the attraction of cheap, reliable computing would be irresistible.
But Carr’s analogy to the power grid can only be taken so far. Electricity is a one of the simplest of commodities. (Basically the producers and users just have to agree on a few standards, like voltage and frequency.) Information production, i.e., computing, is much more diverse, encompassing everything from word processing to video streaming to nuclear weapons simulation. Different types of computing suggest the need for heterogenous computing environments, which many not be practical for a utility datacenter. Also, unlike electricity, computing usually assumes some level of security and privacy — something that is difficult to achieve in mixed-use, off-site facilities.
A better analogy for computing may be food production. When food became a commodity, agribusiness conglomerates took over and replaced lots of family farms with much larger, more efficient “factory farms.” Today, crops like wheat and soybeans are typically grown on multi-hundred acre land parcels. But not all food products are easily commoditized. Specialty fruits, vegetables, and organic products don’t usually lend themselves very well to large-scale production. According to the U.S. Department of Agriculture about a quarter of farm revenue is still generated on family farms. Many of these farms are focusing on these specialty items and have formed cooperative arrangements in order to remain economically viable.
In that sense, high performance computing may be more analogous to an enterprise like a winery. While large-scale wine production is certainly present, a significant number of growers have banded together on a regional basis to produce higher value products: fine varietal wines. In the latter case, individual wineries operate on a much smaller scale than a typical agribusiness enterprise. Something similar is actually taking place in non-commercial HPC, where regional grids, like TeraGrid or EGEE, or more loosely-coupled organizations, like the U.S. DOE labs or Europe’s DEISA supercomputing consortium, have formed alliances to share computing resources.
It may be the case that non-commercial supercomputing will stick with this model, while most commercial high performance computing will eventually end up in the cloud. But I can’t help but think HPC will be one of the last adopters of the utility model. The computing cloud is being built on top of the Internet — a high-latency, low-performance computational infrastructure designed mainly for data storage and distribution, not computation, and certainly not scientific computation. It will be years before even desktop and mainstream enterprise computing are incorporated. The cloud is going to need to develop some serious thunderheads to suck in high performance computing.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].