PNNL's Sriram Krishnamoorthy is one of many scientists reaching for the next step in supercomputer evolution, the exascale computer. DOE has awarded him $2.5 million over five years to explore ways to advance exascale computing through their Early Career Research Program.

The top supercomputers nowadays work at the petascale level, performing in one hour what would take a typical laptop more than roughly 20 years to do. But as computer programs that help solve energy and environmental problems get more useful, they also get much bigger. Exascale computing seeks to solve problems that are about one thousand times bigger than what the top computers can do today.

That magnitude requires supercomputers to perform different parts of calculations simultaneously, sometimes on different kinds of computer hardware, and then put all the pieces back together on the fly. This computational style is called parallel computing and its complexity creates challenges such as making sure all the parts of the system are working as well as they can be. In addition, complex, multi-component calculations have more chances to err and crash. Krishnamoorthy has been studying ways to make computers better deal with these issues.

Currently, computational scientists must translate equations that work on conventional machines into computer language and a style that can be used by a parallel computer. Krishnamoorthy has begun to automate parts of this process. He has also created a set of tools that allows programmers to write code in modules that can be automatically matched to different computing platforms, making it easier to customize programs to different systems.

He has also improved how supercomputers handle errors that could make them crash. When a fault crops up, supercomputers return to the last good checkpoint. By creating programs that identify just the work lost due to a fault and only redo that portion, Krishnamoorthy has narrowed the amount of work that a supercomputer has to repeat. This can greatly improve the speed of science on supercomputers.

He will be using the new support from DOE to delve deeper into how parallel computing solves problems and making sure that the different pieces of the full calculation are working as efficiently as possible. After understanding when and where certain approaches work best in different programs and platforms, he will be testing how they will perform on the computer systems of the future.