Does it make sense to rent CPU hours instead of acquiring computers?
February 15, 2005 | Computers for high-performance computing are relatively cheap and getting cheaper all the time. However, the non-human-resource cost of running the machines (HVAC, electricity, space) is becoming increasingly significant compared to the actual hardware acquisition and maintenance costs — not to mention the hidden cost of periodic infrastructure renovation.
At what point does it make sense to "rent" CPU hours as opposed to acquiring computers and maintaining a facility to support them?
The concept of locating HPC resources at a remote facility was perhaps first shaped by governmental supercomputer centers. Researchers would use their own computers to develop their analysis. When they were satisfied with the analysis algorithm, they would port it to the supercomputer platform (remember UNICOS?) along with their data for production computing. The supercomputer facility could be located anywhere — it did not matter as long as the researchers had a good network connection to it.
The mention of "collocation" conjures up thoughts of hosted Web and e-mail servers. However, commercial collocation facilities (such as Virtual Compute Corp. or Boston Datacenters), in collaboration with major computer vendors, have evolved well beyond this and have made vast computing resources available for scientific research. For the purposes of this column, collocation refers to purchasing access to computing resources located elsewhere.
Why hasn't life science research been drawn to collocation?
One prominent reason is trust. Trust encompasses technical issues such as ironclad secure authenticated and authoritative access to assure the client that his or her computation is protected from intrusion from a competitor sharing the same collocation facility — perhaps even the same collocation hardware. Many of these technical issues have been thoroughly addressed through software and hardware technologies. Clients may also have non-technical trust issues from not having physical contact with the hardware. Regardless, it is the burden of the collocation providers to convey the concept of trust to their clientele.
If it were only about computer hardware, then collocation may make economic sense. If in-house, cluster compute resources are often idle, there is no doubt there are cost savings associated with purchasing computing in units of minutes rather than units of racks. This may come at the price of the minor complexity and inconvenience of remote access. However, for many large research groups the cost of acquiring and maintaining computing hardware is a very small component when compared to the cost of the software that runs on it. The software costs of commercial databases and specialized computer scientific applications overshadow the cost of the required hardware — sometimes as much as 50-fold. At this cost scale, the cost and location (in)convenience of the compute resource become increasing irrelevant.
Expense vs. Asset
Other considerations when debating "expense versus asset" computing are:
- For some research groups, a full account of access to the compute hardware (both physical and electrical) must be maintained to provide an audit trail to preserve intellectual property and research integrity. Collocated access accountability is possible but may be overshadowed by the introduced complexity.
- From a financial accounting perspective, capital assets are budgeted differently than ongoing expenses. Acquiring and maintaining the physical hardware provides the assurance that researchers will have computing resources in the face of unpredictable expense budgets.
- The collocation company may simply go out of business. Commercial collocation is highly competitive, and margins are slim on traditional hosting solutions.
Some argue that internalizing and managing compute resources is a core competency that also helps with developing new innovation and products. Outsourcing this may lead to a reduced innovation. Yet there may be an inflection point where collocated compute access may be highly desirable. This inflection point will depend upon the size of the client company and the problem domain, and the quantity of computing. For a small company seeking vast compute resources for intermittent projects, collocated computing may be the more attractive option.
Michael Athanas, Ph.D., is founding partner of The BioTeam. E-mail: firstname.lastname@example.org.