The PALM Project: MPMD Paradigm for an Oceanic Data Assimilation Software
The PALM project aims to provide a general structure for a modular implementation of a data assimilation system. In this system, an assimilation algorithm is split up into elementary “units” such as the observation operator, the computation of the correla
- PDF / 198,547 Bytes
- 8 Pages / 431 x 666 pts Page_size
- 1 Downloads / 166 Views
Abstract. The PALM project aims to provide a general structure for a modular implementation of a data assimilation system. In this system, an assimilation algorithm is split up into elementary ”units” such as the observation operator, the computation of the correlation matrix of observational errors, the forecast model, etc. This approach allows to separate the physical part of the problem from the algebraic part. PALM performs the algebra, ensures the synchronization of the units and drives the communication of the fields exchanged by the units. Thanks to the interface specifications, each physical unit can be implemented as an independent executable code and can be easily replaced or reutilised for other configurations.
1
Introduction
The MERCATOR project aims to implement an oceanic forecasting system with assimilation of satellite-born and in situ data. This system will be used as a research tool and for operational purposes. This double aim imposes the full modularity of the whole system and the best tuning of the performances on all the actual reference plate-forms. Modularity, performances and portability are three aspects hardly ever compatible: the best trade-off is looked for. The core of a forecast system with data assimilation is the assimilation cycle: this is a large scale optimisation problem which makes use of very different computational components, such as the forecasting model, the observation operator mapping the state of the model onto the data space, the statistics, the minimiser and so on. Each component is best implemented as a stand alone code developed and tested by a separate scientific team. The effort due to the merging of these heterogeneous components in an efficient code is the main hindrance for the research in data assimilation. The philosophy of the PALM software, the back-bone of the MERCATOR project, is to keep the components as independent as possible and to make them work together in a MPMD distributed system. The definition of algorithm independent interfaces will ensure the full modularity of the scheme. This document presents the motivation for such an approach on a simple example and the implications that the users requirements have on the design of P. Amestoy et al. (Eds.): Euro-Par’99, LNCS 1685, pp. 1423–1430, 1999. c Springer-Verlag Berlin Heidelberg 1999
1424
A. Fouilloux and A. Piacentini
the code. The PALM philosophy is explained in detail in the central sections. The technical choices are then sketched in section 5.
2
MPMD for Data Assimilation
Let’s consider as a simple example the computation of the gradient of the cost function J in the 3D-Var assimilation scheme ∇J(x) = P −1 (x − xb ) + HT|x R−1 H(x) − y0 where x is the optimal initial condition to be found by minimization of J, xb is the first guess previously forecast, y0 are the observations, H is the operator going from the model space to the space of the observed variables and P , R are the variance/covariance matrices of the forecast and the observation errors. Some remarks determine the ch
Data Loading...
