Parallelization and performance of a meteorological limited area model
Introduction
In this study, we consider the BOLAM system [1], a research-oriented Limited Area Model (LAM) to study weather events in a complex environment. It is divided in three parts: initialization routines, meteorological model, and post processing routines. Here, the focus is on the meteorological model, the most complex and computationally expensive module. Typical spatial resolutions of state-of-the-art LAMs are in the order of a few tens of kilometers.
However, there is a strong interest in the investigation of their performance, in particular the dynamics, radiation and convection parameterizations, near the limit of validity of the hydrostatic approximation at about 7–10 km. For these reasons, we have developed a parallel version of the model to see if the requirements of higher spatial resolution and more frequent calls of the physical package can be met by the present-day generation of parallel computers. We have used a standard domain decomposition technique and considered a few message-passing libraries to ensure portability. The paper is organized as follows: in Section 2, we briefly reviewed the analytical model and the numerical integration scheme. In Section 3, we discussed the domain decomposition strategy and a few parallel implementation issues, while in Section 4, we presented the results of a set of numerical experiments. We stated our conclusions in Section 5.
Section snippets
Overview of the BOLAM model
This meteorological model integrates in time the set of partial differential equations that constitute the so-called primitive equations model. The prognostic variables are: zonal u and meridional v wind components, surface pressure ps, potential temperature θ, and specific humidity q. The horizontal discretization of the model equations is done on the staggered Arakawa C-grid using centered finite differences, with latitude φ and longitude λ as independent variables. The σ coordinate, defined
Domain decomposition and parallel implementation issues
A meteorological model is usually a collection of very specialized routines, which are often optimized for old generations of vector computers. Consequently, we have redesigned the entire model to port it on a distributed memory parallel computer. The new scalar version is running 30–40% faster. The spatial domain is usually a rectangle in the (φ, λ) plane, and a domain decomposition, with each subdomain assigned to a different processor, is the natural way of dividing the work load. The choice
Analysis of performance
We have conducted a series of experiments to study the performance of the parallel model in the three test cases described in Table 1. We monitor the time spent in computations and communications on each processor. All times are relative to a fixed number of calls of the time-stepping routine. In all comparisons, we do not consider the I/O and the initialization times. The reason is that these times are of different nature in the serial and parallel model and several options are also available.
Conclusion
We have discussed several problems concerning the parallelization and performance of a meteorological model. We have considered several layout of the interprocessor communications and different communication strategies. The results show that for low resolution problems only a very coarse-grain subdivision is meaningful. As the size of the problem increases, better efficiencies are achieved, up to 85% of the theoretical maximum, relatively to the available number of processors. For this test
Acknowledgements
The authors acknowledge the support from the Sardinian Regional Authorities. We had interesting and helpful discussions with R. Benzi, M. Manzini, C. Vittoli and G. Zanetti.
References (16)
- et al.
Accurate predictions of parallel execution time
J. Parallel Distributed Computing
(1995) - et al.
Monitors, messages, and clusters: the p4 parallel programming system
Parallel Computing
(1994) - et al.
Validation of a limited area model in cases of mediterranean cyclogenesis: surface fields and precipitation scores
Meteorol. Atmos. Phys.
(1994) Wind stress on the water surface
Q. J. R. Meteorol. Soc.
(1955)A scheme for representing cumulus convection in large scale models
J. Atmos. Sci.
(1991)- et al.
An economical analytical method for the computation of the interaction between scattered and line absorption of radiation
Contrib. Atmos. Phys.
(1979) - J.F. Louis, M. Tiedtke, J.-F. Geleyn, A short history of operational PBL-parameterization at ECMWF, ECMWF Workshop on...
- et al.
A comprehensive radiation scheme for numerical weather prediction models with potential applications in climate simulations
Mon. Wea. Rev.
(1992)
Cited by (5)
Dataflow Computing in Extreme Performance Conditions
2015, Advances in ComputersCitation Excerpt :The application accelerated is a research-oriented complex atmospheric model developed by ISAC-CNR (Bologna, Italy) [8]. The application falls in the category of a hydrostatic limited-area model (LAM) parallelized using domain decomposition and message passing libraries [9]. Historically hydrostatic models present a considerable computational advantage over their nonhydrostatic counterpart.
Multifractal analysis and simulation of rainfall fields in space
1999, Physics and Chemistry of the Earth, Part B: Hydrology, Oceans and AtmosphereAcceleration of a meteorological limited area model with dataflow engines
2012, Symposium on Application Accelerators in High-Performance ComputingAn economical second-order advection scheme for numerical weather prediction
1999, Quarterly Journal of the Royal Meteorological SocietyMultifractal modeling of anomalous scaling laws in rainfall
1999, Water Resources Research