This winter, JCSDA’s Francois Hebert completed a code replacement resulting in significant impact: updating the interpolation between the GSI-B regular grid and the JEDI model grid in the SABER wrapper. SABER (System Agnostic Background Error Representation) is JEDI's collection of algorithms that model errors in the forecasts and background states. GSI-B (Gridpoint Statistical Interpolation–Background) is the SABER model that interfaces with the GSI data assimilation system, used by our partners at NASA and NOAA.

The old code was inefficient in parallel environments but remained heavily used by our partners at NASA and NOAA running experiments with GSI-B inside SABER. This highlights the key “joint” aspect of JCSDA; with feedback from in-kinds and other users at these agencies, the JCSDA core team, as the code owners, identified the slow run times linked to this particular interpolator and replaced it entirely with a new one.

The results are impressive. The new interpolator hasn’t just made the code slightly more efficient - it has quadrupled the run speed for a 3DVar experiment with GSI static B. Ricardo Todling at NASA GMAO then tested it on a hybrid 4DEnVar experiment with GSI hybrid B and full observing system, achieving equally good results (see figure 1).

This collaboration highlights the essential roles of both the core team and our partners in enhancing JEDI’s operational readiness.

Figure 1. Blue bars show the run time for the develop code on 192 and 672 processors, and purple is the runtime for the new SABER code.

Photo by NASA on UnSplash