Speaker
Description
PALM is an open-source large-eddy simulation model with strong focus on parallelization and HPC computing. The members of the atmospheric modelling group at the Institute of Computer Science (ICS) of the Czech Academy of Sciences successfully extended this model with the Urban Surface Module (USM) and Radiative Transfer Module (RTM), which simulate processes necessary for microscale urban climate modelling (3-D multi-reflective radiation, surface energy balance, 3-D resolved vegetation, basic air quality modelling and others). The extended model has been verified in a pilot study [1]. The PALM-4U [2] is a new complex urban climate modelling system developed by multiple research institutions from Germany, Czech Republic and other European countries, which is built on PALM-USM for urban layer radiation and surface energy processing. The ICS team is a core member of the PALM-4U development team.
Modelling multi-reflective radiation within fully resolved 3-D terrain, buildings and vegetation is a complex task which brings many issues that are not normally encountered in atmospheric modelling. One of the crucial problems is the fact that in a highly parallelized atmospheric model, each process simulates only a small part (subdomain) of the whole modelled area, it has no direct access to values outside it and a typical data exchange pattern involves only neighbouring subdomains. With 3-D radiation, there are direct interactions among arbitrarily distant surfaces and vegetation. Another significant problem presents determining mutual visibility between individual surface and plant canopy elements.
The Radiation Transfer Model (RTM) within PALM-4U uses specially designed raytracing algorithm to establish information about mutual visibility. This differs significantly from a typical raytracing algorithm used in computer graphics due to different data structures and parallelism used in atmospheric modelling. The raytracing process, although highly optimized, is computationally demanding and data-transfer intensive, therefore it is performed as part of model initialization and its results are stored in the form of view factors. During model time-stepping, the precalcuated view factors are used to calculate the actual radiative processes with much smaller demands on CPU time, memory size and data transfers.
We will present the basic principles of the RTM and its implementation and demonstrate its performance, efficiency of parallelization and its scalability on dedicated performance testing simulations.
