LLNL Team Wins SC21 Reproducibility Advancement Award for Approximation Framework

NS. LOUIS, Nov. 18, 2021 — A suite developed by a team at Lawrence Livermore National Laboratory (LLNL) to simplify the evaluation of approximation techniques for scientific applications has won the first-ever Best Reproducibility Advancement Award at the 2021 International Conference for High Performance Computing, Networking , storage and analysis (SC21).

The award, recently established by the conference, recognizes outstanding efforts in promoting the transparency and reproducibility of high performance computing (HPC) methods. The conference organizers presented the inaugural award to the LLNL researchers for a paper describing their High Performance Approximate Computing (HPAC) framework, which allows users to easily evaluate the accuracy and performance tradeoffs of various approximation techniques on OpenMP HPC applications. to investigate.

“We are honored to receive the first-ever SC Best Reproducibility Advancement Award,” said lead researcher Harshitha Menon, a computer scientist at LLNL’s Center for Applied Scientific Computing. “SC has been a pioneer in reproducibility, so it’s very exciting to get an award in this area from SC. It speaks to the importance SC attaches to reproducibility and transparency when reproducing results.”

As the computer reaches the limits of Moore’s Law — the concept that processing speed doubles about every two years — researchers are exploring new paradigms for improving performance in HPC. While approximate calculations — techniques that produce results that are “nearly correct” — can significantly improve performance, adoption for scientific applications is limited due to strict accuracy requirements, Menon explained.

Scientists’ uncertainty and reluctance to rely on the final results requires a better understanding of the approximate performance and accuracy tradeoff and ensuring that methods used for evaluation are reproducible, researchers said.

“Estimated computing has been studied for many years and has gained some traction over the past decade, but most of the current solutions are more of an ad hoc solution, and this has led to implementations that are missing,” said lead author and LLNL postdoctoral researcher Konstantinos (Dinos) Parasyris. “In our case, we tried to use state-of-the-art tools and provided software that can be used by someone else.”

In designing HPAC, the LLNL team integrated advanced approximation techniques (loop punching, input/output memory), with a common LLVM/Clang compiler and OpenMP runtime support. By staying close to the OpenMP interface, HPAC makes it easy for users to specify which technique they want to use, annotate some of the code to try it out and get a first idea of ​​performance improvements, researchers said. Users can compare different approach techniques and, depending on their application and fault tolerance, make an informed choice as to whether techniques are useful to them or not, Menon said.

“The framework provides a way to generate the plots and say, ‘this is the range you get for these particular approximation techniques, and now it’s up to you to choose the appropriate error range for your application and maximum speed,'” Menon said. . “There’s a game of accuracy versus performance trade-off that we’re playing here, so depending on what the user needs, they can go one way or another. Usually there’s a sweet spot.”

In the paper, the team applied HPAC to eight commonly used HPC benchmarks, and found that the suite provided significant performance gains for certain error thresholds. For example, in the LULESH benchmark – which approximates the hydrodynamic equations – they found that approximation yields significant performance gains of up to 1.7 times, due to the reduction in memory requirement. In an application called leukocytes — which detects and tracks white blood cells in video microscopy of blood vessels — the team found that, in addition to increasing speed by increasing the number of threads, HPAC offered an additional speed of up to 25x at an error threshold of less than 5 percent. .

HPAC is funded by the Laboratory Directed Research and Development program and is part of an LLNL “ApproxHPC” project that aims to evaluate approximation techniques and create tools to increase confidence in estimated computing for scientific applications in the post- Moore’s Law era.

The team continues the work by further exploring techniques and improving the understanding of error susceptibility. The HPAC software is available on GitHub. Co-authors include LLNL scientists Giorgis Georgakoudis, James Diffenderfer, Ignacio Laguna, Daniel Osei-Kuffuor, and Markus Schordan.

Source: LLNL

Leave a Comment