| Home > Research > Ifsdocs > WAVES > |
Chapter 2. The kinematic part of the energy balance equation
Chapter 3. Parametrization of source terms and the energy balance in a growing wind sea
Chapter 4. An optimal interpolation scheme for assimilating altimeter data into the WAM model
Chapter 5 Numerical scheme
Chapter 6 The WAM-model software
Chapter 7 Wind-wave interaction at ECMWF
REFERENCES
Previous Section
6.2 Massive Parallel Computing
Memory distributed machines such as the Fujitsu vpp series require the introduction of message passing between processors (known as processing elements (PE's)). Therefore, one PE can send a message which is received by one or more other PE's. In its very basic implementation, the message is nothing more than a one-dimensional array of a given type containing values that are needed by the other PE(s) plus the necessary information about the sender and receiver. For a succesful message exchange both send and receive should be completed.
Message passing in the ECMWF version of the WAM model was introduced in 1996 based on the message passing library MPELIB. This newly developed code can also run on non distributed memory machines and on a single PE.
In the present setting the number of processors is determined at run time. Once the message passing program starts simultaneously on all assigned PE's, the parallel environment and the message passing protocol is initialized. Also, the total number of PE's is determined as well as the logical PE on which the code is run. This initialisation procedure is done in CHIEF, or any other start up main program which calls the WAM model.
When running in parallel it is important to have an even distribution of work over the PE's in order to avoid load unbalance. In the case of the WAM model it comes down to splitting the global computation domain into regions of equal size, keeping in mind that information is only locally known on each PE and can only be exchanged with the other PE's via message passing(which is a slower process than computing).
Once the total number of PE's is known, MPDECOMP is called to set up an even decomposition of the total grid into one sub domain per PE. Since the global grid is mapped onto a one-dimensional sea point array following increasing latitude lines, the sub domains are chosen to be consecutive segments of the full sea point array (cf. Fig. 6.2 ).
The length of each segment is determined by the requirement that the work is distributed in an even manner over the given number of PE's. Thus, each PE will only perform the integration of the source functions of the energy balance equation of one subdomain. However, the upwind scheme which solves the advection term, uses neighbouring grid points in the 2-D grid that might belong to another subdomain. The necessary information from the other PE's needed to evaluate the spatial derivatives of the energy flux are obtained through message passing. Here, the message is constructed using the geometrical rules displayed in Fig. 6.2 , and is similar in spirit as the method that was developed for the multi-block version of the WAM model. An important difference is, however, that the domain composition is done at run time, allowing more flexibility.
This completes our discussion regarding the design of the WAM model. The remainder of this subsection is devoted to the model system. A more detailed description of this may be found in the manual wamodel cycle 4 by Günther et al. (1991).
The model system consists of three parts:
| 1) pre-processing programs, |
| 2) processing programs, |
| 3) post-processing programs. |
The original WAM model is designed to run as a module of a more general system or as a stand-alone program. The pre-processing programs generate the model grid, bathymetry-dependent dispersion relation, etc. Post-processing programs are provided for archiving and for further analysis of the model output.
Next Section
Previous Section
 |
23.04.2002 |
 |
© ECMWF |
