Supercomputing: New Processor Architecture Holds Promise for Protein, Gene Studies
A supercomputing architecture that first appeared in prototype form more than 10 years ago has been given a new lease on life, thanks in part to a recent $4 million Department of Defense grant issued to seed the new Center for Adaptive Supercomputing Software. The joint project teams up Pacific Northwest National Laboratory and supercomputer maker Cray, as well as several institutions including Georgia Institute of Technology and Sandia National Laboratories.
The initiative aims to develop software that takes advantage of the multithreaded processing capabilities of Cray’s XMT supercomputer. Unlike traditional supercomputer processing architecture, where each processor gets a portion of memory for each calculation in a piece-by-piece fashion, the new processors are each capable of multiple, simultaneous data crunching and use a much larger pool of memory per processing core. This design means that many disparate sets of complex data can be digested at once, instead of each portion of data being handled piece by piece.
David Bader, a computer scientist at Georgia Tech, demonstrated the architecture’s application to biology by identifying proteins that, when knocked out, disrupt the cancer-causing networks in a particular cell. Bader and his team used a social networking algorithm to mine a huge collection of publicly available human proteome datasets. “This is similar to finding important people in a social network, sometimes called ‘connectors,’” Bader says. “Looking for these proteins is like looking for a needle in a haystack — and it is usually computationally intensive that won’t work well on current [high-performance] machines, but this new architecture is really designed for this type of problem.”
Normally, multiple database searches of this kind would take hours and hours to complete on a typical cluster or supercomputer. But with an algorithm specially ported to this multithreaded processing architecture, the same job takes mere seconds to complete, says Bader. “These sorts of problems have overwhelmed modest size clusters, and if you start adding processors to a cluster, it takes longer and longer to run because the communication costs dominate,” he says. “This is really the first architecture where you can pose a biological hypothesis, test it out, and run it in short seconds or minutes versus hours to days, or maybe never.”
Bader and his colleagues believe the concept could offer a lot to largescale life science computing problems. “I think as we gather more genomics data we can use such a system to make scientific discoveries,” he says. “I would hope, looking at three to five years, that we keep investing in these novel types of architecture and looking at the scientific results we can achieve, especially in the areas of solving genomic problems and understanding of the genome.”
— Matthew Dublin