see NASA's 4 dimensional Data Assimilation Grand Challenge for more details of Makivic analysis of HPF for this application
|
|
Van Leer and Prather methods are monotonic finite difference schemes for fluid advection which have good behaviour regarding diffusive and phase errors
|
Very suitable for HPF implementation:
-
regular grid, uniform decomposition
-
structured communication (shifts, scans)
-
computation adequately expressed using array syntax, FORALL and HPF library
|
Load imbalance due to polar subcycling eliminated via gather/scatter procedure, which takes just a few code lines using array subsection notation (would have been tedious work in message passing!):
-
polar regions gathered into a temporary array
-
temporary array mapped onto the whole machine
-
advection performed on the temporary array
-
temporary array scattered back into the original array
|
|
Communication costs due to gather/scatter are smaller than load imbalance costs
|
|
Big advantage of HPF implementation: very easy to experiment with different decomposition strategies
|
|
For example, depending on hardware parameters and/or model grid resolution either two-dimensional or one-dimensional decomposition may be optimal.
|
|
Simple <em>preprocessor directives</em> are sufficient to implement either decomposition using HPF: <em>bulk</em> of the code would have to be changed to go from one decomposition to another using message passing.
|
|
Performance: 2.5 GFLOPS sustained and 6.8 GFLOPS peak on a 256 node CM-5 for a 144 X 88 latitude/longitude grid. Much better performance can be achieved on finer grids.
|
