The estimated cluster efficiency is shown in
Figure 3. Three resolutions were used for model evaluation:
,
, and
. The largest size corresponds to the
production code currently used in the ECMWF. The smallest case is just the lower
limit where parallel runs on a C90 cluster might be useful. For all these cases
we obtained a good speed-up. The two large examples show an efficiency between 80
and 90%. Therefore, a C90 cluster of P systems is able to execute the IFS for
large problem sizes nearly P times faster than a single system.
Hence our main result is that the parallel IFS runs efficiently on C90/Hippi clusters for sufficiently large examples.