The communication overhead in a parallel implementation of a numerical weather prediction model may possibly partly be hidden behind calculation time. We tried this in a proof-of-concept study with the HIRLAM Eulerian explicit grid point dynamics. We found that savings in total execution time depend on the available hardware. If the network is fast, the additional calculation costs are not compensated by the reduction in communication overhead. But if the interconnection network is not very fast, or on a Grid computer, we are able to almost completely hide communication time. We anticipate several problems before our results can be implemented in an operational configuration of HIRLAM, not the least of them being the code complexity, in particular in a semi-Lagrangian scheme.
[1]
Ian Foster,et al.
The Grid 2 - Blueprint for a New Computing Infrastructure, Second Edition
,
1998,
The Grid 2, 2nd Edition.
[2]
Lex Wolters,et al.
A dynamic application-driven data communication strategy
,
2004,
ICS '04.
[3]
Lex Wolters,et al.
Tomorrow's weather forecast: automatic code generation for atmospheric modeling
,
1997
.