Computer scientist Ignacio Laguna is part of a team of LLNL researchers awarded the prestigious Hans Meuer Award. The award honors the most outstanding research paper submitted to the International Supercomputing Conference High Performance Conference (ISC23) in Hamburg, Germany on May 22, 2023.
The team was recognized for their paper, “Expression Isolation of Compiler-Induced Numerical Inconsistencies in Heterogeneous Code,” which introduces a tool the collaborators developed for identifying the sources of inconsistencies in a code’s calculations. The tool, called Compiler-induced Inconsistency Expression Locator (CIEL), helps scientists efficiently find and fix the root cause of numerical disparities.
When programmers compile their code for a given application, optimizing or updating the compiler they use can affect how the code runs, “which creates a problem for scientists,” Laguna says. This may result in different outputs for the same computation, affecting the reliability of the calculation and the reproducibility of scientific results. These are the discrepancies CIEL aims to find, with applications ranging from weather forecasting to hydrodynamics energy calculations.
Though this effect is well-known within the precision computing community, the search for a solution has been a manual effort in the past.
Simulations run at the Lab can require up to millions of lines of code. With such large applications, it’s difficult—if not impossible—to find which specific line is causing a problem by hand.
“It’s like if I give you a book with thousands of words, and you know there’s a problem in one chapter’s sentence,” explains Laguna, who leads the Parallel Systems Group in LLNL’s Center for Applied Scientific Computing. “How do you go and find that?”
To date, CIEL is the only automated tool that can isolate these inconsistencies. And it does so very efficiently: CIEL can locate the source of the disparities with over 99% precision.
This is Laguna’s second time winning the Hans Meuer Award. He states, “It’s very exciting. Sometimes, you’re lucky enough to submit a paper that the reviewers like a lot, but that’s usually once in a lifetime. This is the second time, so it’s surprising and I’m very grateful.”
Laguna shares the award with his co-authors, Dolores Miao and Cindy Rubio-González, from the University of California, Davis. CIEL’s source code is publicly available.