Regional atmospheric models are important tools for short-range weather predictions and future climate change assessment. The further enhancement of spatial resolution and development of physical parameterizations in these models need the effective implementation of the program code on multiprocessor systems. However, nowadays typical cluster systems tend to grow into very huge machines with over petaflop performance, while individual computing node design stays almost unchanged, and growth is achieved simply by using more and more nodes, rather than increasing individual node performance and keeping adequate power consuming. This leads to worse scalability of data-intensive applications due to increasing time consumption for data passing via clusters interconnect. Especially some of numerical algorithms (e.g. Those solving the Poisson equation) satisfactorily scaling at previous generation cluster systems do not utilize the computational resources of clusters with thousands cores effectively. This prompts to study the performance of numerical schemes of regional atmospheric models on processor architectures significantly different from those used in conventional clusters. Our approach focuses on improving the performance of time explicit numerical schemes for Reynolds-averaged equations of atmospheric hydrodynamics and thermodynamics by parallelization on CellBE processors. The optimization of loops for numerical schemes with local data dependence pattern and with independent iterations is presented. Cell-specific workloading managers are built on top of existing numerical schemes implementations, conserving the original source code layout and bringing high speed-ups over serial version on QS22 blade server. Intercomparison between Cell and other multicore architectures is also provided. Target ing the next generation of MPI-CellBE hybrid cluster architectures, out method aims to provide additional scalability to MPI-based codes of atmospheric models and related applications

Ключевые слова: cell broadband engine, MPI clusters, atmospheric modeling