|Name:||Architectures(8) Source-to-Source Translation for Climate Models|
|Time:||Monday, June 17, 2013
2:00 PM - 2:05 PM
|Room:||Multi-Purpose Area 4 (MPA 4)
CCL - Congress Center Leipzig
|Speakers:||Raul Torres, University of Hamburg|
|Abstract:||The complexity of the ICON climate model and the presence of machine-specific structures inside it, make the code overall extensively complex and inflexible. For these cases, we can use abstraction of data and loop structures that allow a more natural description of the model and also flexibility for adapting to architectures. Currently we extend Fortran syntax to allow such abstraction. By providing a Domain Specific Language (DSL) and a source-to-source translator we can ease the programming task in the climate scientist side. On the one hand, the DSL acts as an extension to Fortran, with the ability of simplifying the expression of some operations and the detaching of the language from a particular hardware platform. The translator, on the other hand, converts the DSL code to Fortran, lowering down to the correct selection of instructions according to the machine. We have used ANTLR Parser Generator to build our translator. In a first step, the translator needs to parse a machine-dependent configuration file, where the particular details of the platform are specified. Afterwards, it parses the DSL annotated code and converts it to an intermediate representation called Abstract Syntax Tree (AST). From this structure, Fortran code is generated back. The DSL abstractions were applied on the ICON testbed and executed on a Power6 cluster architecture. The results showed that the number of iterations per second was raised more than 10% when 32 cores or more were used, with a maximum of 17% when 64 cores were used.
Julian M. Kunkel, DKRZ; Leonidas Linardakis, MPI for Meteorology; Raul Torres, University of Hamburg