Google is chasing Amazon Web Services in a new area –- providing access to “on-demand” supercomputing resources.
On Wednesday, Google announced the general availability of something called pre-emptible virtual machines, which amount to computing resources you can rent for very little. The service, firstannounced in May, enables people to buy computer processing that Google isn’t using at a steep discount, provided the customer is willing to yield the resources on short notice.
A.W.S. has a similar short-term usage program, called Spot Instances. It lends itself to supercomputing exercises, particularly for universities and companies that don’t want to buy multimillion-dollar machines of their own, and has been used in areas like drug discovery and learning about new materials.
The addition of Google as a competitor is likely to increase the number of ways these resources are used while lowering prices.
Google also appears to be stressing its skills in machine learning, a powerful tool for finding new patterns in large data sets.
Using Google, cancer researchers at the Broad Institute used 51,200 computing cores to look at the interrelationships between human genes, the billions of ways they are expressed, the cell lines from some 500 types of cancer, and molecules that perturb those cells. The idea was to sort through billions of data points quickly, looking for promising areas for researchers to seek drugs and treatments.
The analysis, which on a single computer server would have run about 30 years, took a couple of hours, said Chris Dwan, the acting director of Information Technology at Broad. It cost about $4,000.
“This isn’t like computing a few taxi trips,” said Mr. Dwan. “This is really heavy computing in complex biological systems, steering research.”
He added, “the rules of how we work are being completely rewritten.”
Losing the machines for a short time in these supercomputing exercises is not lethal, as long as the project is designed to back up data and restart. Frequently, the number-crunching in such projects is relatively brief, while sorting out what the results mean can take months.
If Google is serious about offering an alternative to A.W.S., it still has a way to go. Cycle Computing, the company that specializes in cloud supercomputing and set up the project for Broad, has done projects on A.W.S. that use 150,000 cores.
In April, A.W.S. bought a company called ClusterK, which does work similar to Cycle, and will likely become more aggressive in helping customers learn how to use this service.
Jason Stowe, the chief executive of Cycle, noted that running a supercomputing task, as Broad did, costs perhaps one-third as much as it did in 2012. The more interesting advance is in the increasing use of machine learning to steer actual research.
“What happens in Life Sciences, where they have used lots of computers for a long time, tends to lead what happens elsewhere,” he said. “The first time we started doing cloud supercomputing it was drug discovery, then financial services, then materials science. Now it’s not just analysis, it’s directing future decisions.”