IFS Documentation front page

I Observations
II Assimilation
III Dynamics
IV Physics
V Ensemble
VI Technical
VII Waves

4.4 Analysis of CPU time

In developing a high-resolution spectral model, the cost of the transforms (particularly the Legendre transforms) may be a cause for concern (e.g. Côté and Staniforth, 1990). In the case of a semi-Lagrangian model, it is clearly important that the gain obtained through the use of longer time steps is not outweighed by the extra cost of the semi-Lagrangian scheme. In view of these concerns, it is of interest to analyse the CPU time required for our model. Table 4.1 shows the percentage breakdown for the Eulerian version, for the fully interpolating semi-Lagrangian scheme and for the vertically non-interpolating scheme, at T213/L31 resolution.

Table 4.1 Analysis of CPU time (%)

	Eulerian	Fully interpolating semi-Lagrangian	Vertically non-interpolating semi-Lagrangian
Dynamics	21	15	17
Physics	53	42	45
FFT	6	3	4
Legendre transforms	20	13	14
Semi-Lagrangian		27	20

This analysis suggests that the spectral method is still perfectly viable at this resolution, and that considerably higher resolutions can be achieved before the cost of the transforms becomes a matter for serious concern. The overhead of the semi-Lagrangian scheme, particularly the non-interpolating version, is also quite modest; for the present resolution it permits a timestep of 15 minutes compared with 3 minutes for the Eulerian version, and the resulting reduction in the CPU time for the forecast is about a factor of four. The semi-Lagrangian overhead is in fact slightly less than suggested by the figures in Table , since there is a simultaneous reduction in the number of transforms compared with the Eulerian scheme. Comparing the two variants of the semi-Lagrangian scheme, the overall CPU time for the non-interpolating version is 8.5% less than that for the fully interpolating version.