Jan 20, 2021
Scientists around the world and at GSL are working on fundamental changes to computer code to improve the computational efficiency of weather models. In early December 2020, four GSL researchers participated in intensive “hacking” sessions with mentors from NVIDIA and Lawrence Livermore National Labs to accelerate their computer code towards an exascale-ready model that encapsulates numerics, science, and software engineering.
NVIDIA and NOAA hosted the “hackathon” virtually and used the latest NVIDIA technology and techniques. During “scrum” sessions, a forum in which coders can address complex software problems, hackathon mentors worked with GSL researchers on specific algorithms within GSL’s GeoFLuid Object Workbench (GeoFLOW). GeoFLOW is a framework for partial differential equation solvers to enable the rapid development of computational models relevant to problems in weather forecasting and atmospheric science. Only a portion of the code was targeted for the hackathon.
The team successfully enabled the targeted code to operate not only on Central Processing Units (CPUs) but more importantly on advanced Graphical Processing Units (GPUs). GPUs are able to solve math problems and do calculations more efficiently using massive levels of parallelism that may drastically speed up the time to a solution -- good news for weather models.
GSL Bryan Flynt (CIRA) led the effort by preparing the submission and documenting GSL’s GEOFLOW code in detail so the mentoring team could download, build, and run it in real-time. Bryan also led and presented all hackathon group presentations. Duane Rosenberg (CIRA) designed and developed the target code and driver. Daniel Abdi (CIRES), and Jacques Middlecoff (CIRA) also participated.