| Home > Research > Ifsdocs > PHYSICS > |
DYNAMICS
IFS documentation Front Page
Table of contents
Chapter 1. Introduction
Chapter 2. Basic equations and discretization
Chapter 3. Semi-Lagrangian formulation
Chapter 4. Computational details
Previous Section
4.2 Multitasking
Currently the model is run on a `modestly parallel' supercomputer (specifically, a Cray Y-MP C90 with 16 processors), and multitasking is an important aspect of the strategy to make the best use of the available computer power. We have chosen to rely mainly on high-level `macrotasking', i.e., dividing the computation into large independent units of work, each of which is assigned to one of the processors. Here only a brief outline will be given; additional details and discussion are provided by Dent (1992).
In the first scan, the unit of work is a pair of latitude rows. Each pair is independent of all the others, and a simple dynamic scheduling technique can be used: as each processor becomes free, a new pair of rows is assigned to it.
In the second scan, the unit of work is a single latitude row. For the semi-Lagrangian version, the calculations for each row are no longer independent of those for all the other rows. The trajectory calculations and semi-Lagrangian advection algorithm for the central row of the rotating buffer can only be carried out once the required calculations have been completed for all the neighbouring rows, and somewhat complex logic is required to control the multitasking during this scan.
In the third scan, the unit of work is a single zonal wavenumber. Each wavenumber is independent of all the others, and the scheduling technique used in the first scan can again be used. The work content of each wavenumber varies from a maximum at
to a minimum at the largest value of 
(the `tip' of the triangular truncation), and the dynamic scheduling
technique is effective in spreading the work over available processors.
Next Section
Previous Section
 |
08.04.2002 |
 |
© ECMWF |
