see NASA's 4 dimensional Data Assimilation Grand Challenge for more details of Makivic analysis of HPF for this application
|
Tasks in the operational Data Assimilation system which can be efficiently accomplished using HPF:
-
Search for data within a mini-volume
-
Interpolation of model data onto observation locations
-
quality control: buddy check
-
Quality control: gross check
-
sorting of data within a mini-volume
-
selection of data within a mini-volume
-
assembly of mini-volume matrices
|
|
These tasks are generic enough to be used in both Mini-Volume OI and PSAS
|
There are three distributed templates which correspond to:
-
observational data
-
model grid
-
mini-volume grid
|
|
Data parallel operations which involve a single template usually require structured communication (scans, reductions) or no communication at all
|
|
Data parallel operations which involve mappings between templates require general purpose communications and command highest overhead
|
|
Dynamic memory allocation, reductions, scans and FORALL construct essential
|
|
Independent calculations on mini-volume matrices can be handled via $HPF DO INDEPENDENT or as task parallel computations using $HPF EXTRINSIC facility (this approach can accomodate sophisticated load-balancing schemes)
|
|
Performance: quality control routine for sea-level analysis runs at 0.5 GFLOPS sustained and 3.2 GFLOPS peak for unoptimized CM Fortan code and a small test data set (which cannot use effectively 1024 vector units on a 256 node CM-5). Much better performance can be achieved on production data sets.
|
