The implementation of a global nonhydrostatic anelastic model on the sphere in a lat-lon-z grid is presented. The representation of the orography is realized by cut cells, which describe approximately the intersection of the orography boundary with the grid cells. Therefore the free fluid part in each grid cell is characterized by the free cell volume and the free face area of the six cell faces. Since the lat-lon grid is locally orthogonal the discretization in space can be expressed with the above defined cell characteristics.

The spatial discretized system is solved adaptive in time by a combination of a Rosenbrock-method for the advcetion-diffusion part and a Chorin-type projection method for the pressure. This implicit time integration procedure avoids time step restrictions which are caused by small volume cells at the poles and at the cutting boundary and by fast but unimportant physical processes. The resulting linear systems are solved by preconditioned CG-like methods. For the advection-diffusion part the BICGStab method is used. The preconditioner is built up by an approximate matrix factorization of the transport terms (advection, diffusion) and the source terms (Coriolis, curvature, buoyancy).

The preconditioner for the pressure equation is a multigrid method with a plane smoother to overcome the anisotropies in the grid due to the large horizontal/vertical grid ratio and the pole singularity.

The code is parallelized by a decomposition of the computational domain in rectangular blocks in horizontal as in vertical direction. These blocks are distributed than on the available processors. The decomposition allows also to apply different grid resolution in the individual blocks. One possible application is the use of a coarser resolution in the polar regions. The parallelization requires changes in the solvers for the linear systems. Especially for the pressure computation we use an extended pressure equation with additional flux variables (pressure gradients) at the boundaries between neighbouring blocks. We will present computational results for simplified test scenarios.