Study And Implementation Of Parallel Computing For Three-Dimension Variational Assimilation

Meteorologists have been developing the numerical weather prediction (NWP) technology for accuracy since the early 1950s. In NWP, many data assimilation techniques have been developed to make a good analysis over regions where conventional observations are sparse. The three-Dimensional Variational Assimilation (3DVAR) is an objective analysis approach developed in the late 1980s. Currently, it is one of the most common data assimilation methods used in NWP.As 3DVAR is a problem of high computable and data intensive, parallel computing is the key to take 3DVAR into operation. This thesis presents a simple and fast parallel computing scheme of the Global and Regional Assimilation Prediction System (GRAPES), evaluating the parallel computing performance with respect to complexity, scalability, load balancing, speedup and efficiency. The testing results on a 16-CPU cluster show that the parallel computing scheme achieves considerable load balancing and parallel efficiency. The thesis also reveals the reasons which affect the scalability of the parallel computing scheme through both static and dynamic performance anlysis.The work presented here has the following claims of innovation: (1) it successfully uses the phase parallel model into the parallel computing of 3DVAR, the parallel computing scheme is simple but efficient, (2) it puts forward a new 3DVAR parallel computing scheme with the combination of physics field-parallelism and vertical level-decomposition.