Component concurrency increases the parallel efficiency of Earth system models

Our current model parallelization techniques, such as domain decomposition and loop-level parallelization, decompose the data-space, typically by decomposing the 2-dimensional grid. This limits the availability for parallelism to the availability of enough grid points. An exception to this approach is when running the coupled model, where the atmosphere and the ocean models run in parallel. This is an example of coarse-grained concurrency, where large parts of the model run in parallel throughout a time-step, and are coupled every one or more time-steps.

Component concurrency decomposes the algorithmic space, providing an additional parallelism dimension to the data-decomposition techniques. This increases the parallel workload, that is the amount of operations available within any parallel region. In turn, it allows us to use more computational resources efficiently. Moreover, data-parallel approaches do not help when introducing more processes and details of the Earth system model, such as the cryosphere, atmospheric chemistry, and particle microphysics, while concurrency can absorb part of the additional computational cost. An application to the HAMOCC ocean biogeochemistry model, running concurrently to the ICON-O ocean model, shows that we can double the parallel efficiency on the high-performance computer “Levante” at the German Climate Computing Center (DKRZ).