Sack, Gropp win best paper at PPoPP'12

In supercomputing, existing algorithms for collective communications operations can have limited flexibility and performance. However, University of Illinois Visiting Researcher Paul Sack has come up with new methods for faster communication, and these methods do not require support from additional, expensive hardware that companies might use.

Paul Sack

Earlier this year, Sack and CSL Professor William Gropp (Computer Science) won the Best Paper Award at the 17th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP’12) for their paper about these novel algorithms.

In supercomputing, hundreds of thousands of computers will exchange information to solve a problem. They can communicate through either point-to-point messages, in which one computer exchanges messages directly with one other computer, or through collective operations, in which all of the computers exchange information together to solve a problem.

“I came up with some new methods of doing this that are faster … than any other method of do this,” Sack said. These algorithms take advantage of the topology -- the physical layout -- of a supercomputer and involve sending small redundant messages. This is a pure software approach that does not require any special networking hardware.

“One alternative to my software is companies will add special support for this in the hardware which is very expensive … It’s extra stuff that they have to test and develop and support,” Sack said. “This is kind of a freebie of how to get faster communications without investing in special networking hardware."

Sack originally did all of his work for this project on an IBM system, but said he was able to change just a few lines of his code so that it would be compatible with the University of Illinois’ Blue Waters supercomputer, making his algorithms easily configurable. “That’s a very big thing because … I wrote the code once and I changed a few simple lines of code and it runs anywhere," he said.

Sack said the paper received a lot of attention at PPoPP’12 because no one had conducted research in this area for some time. Also, compared with other work in parallel computing, this was something where the idea was very simple and the results and impact were very large, Sack said.

Sack graduated last December with a Ph.D. in computer science from the University of Illinois at Urbana-Champaign, but has continued his overall research at the University. For more than three years, he has been working full-time as a software engineer at InstaRecon, a company run by CSL Professor Yoram Bresler.