BOB  "Built On Beowulf" Dynamical Core

Some important BOB information

Click here to view t85 3-D simulation of the vorticity field for primitive equations version of BOB (pebob).
 

Click here to view t85 2-D simulation of the vorticity field for shallow water equations version of BOB (swbob).

Download tar-ball for shallow water equations version of BOB model (swbob)

Download tar-ball for primitive equations version of BOB model (pebob)

BOB Documentation

BOB Papers 
Final Bob Paper
Monthly Weather Review

BOB is an efficient spectral dynamical core, written in F77, and primarily designed to run efficiently on small clusters composed of distributed memory computers using MPI. A common problem encountered on such clusters is the limited cache and memory sizes. To overcome this limitation, BOB makes use of several optimization techniques.  First, a recursion relation is used to calculate the associated Legendre polynomials (ALPs) at each timestep ``on the fly'', thus eliminating the need for the order N^3 storage encountered when loading the ALPs from memory.  Second, the need to store the scalar derivatives of the ALPs is avoided by taking advantage of a simple relation between these and the ALPs themselves. The implementation of the Legendre transform is cache blocked in latitude to ensure that the working arrays fit in the cache of each processor, and is unrolled to keep memory traffic low. The cache blocking enables the high performance of the model to be maintained even at very high resolutions. Finally, a one-dimensional domain decomposition and transposition procedure is used in latitude and wave-number space that ensures efficient load balancing among processors.

BOB can be run in two dynamical core versions: either as a dry, adiabatic primitive equations model in pressure coordinates, or as a shallow water equations model. Each version of BOB is downloadable as separate tar files.

BOB was written at NCAR and is the result of a collaboration between Dr. Richard Loft (loft@ucar.edu) of the Computational Science Section in NCAR's Scientific Computing Division, and Drs. Leonard Rivier, Richard Scott and Lorenzo Polvani in the Department of Applied Physics and Applied Mathematics at Columbia University.
 

This page was last updated 04/05/2004