![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Improving Weather Forecasts Through Reduced Precision Data Assimilation |
VerfasserIn |
Samuel Hatfield, Peter Duben, Tim Palmer |
Konferenz |
EGU General Assembly 2017
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 19 (2017) |
Datensatznummer |
250137736
|
Publikation (Nr.) |
EGU/EGU2017-543.pdf |
|
|
|
Zusammenfassung |
We present a new approach for improving the efficiency of data assimilation, by trading
numerical precision for computational speed. Future supercomputers will allow a greater
choice of precision, so that models can use a level of precision that is commensurate with
the model uncertainty. Previous studies have already indicated that the quality of
climate and weather forecasts is not significantly degraded when using a precision
less than double precision [1,2], but so far these studies have not considered data
assimilation.
Data assimilation is inherently uncertain due to the use of relatively long assimilation
windows, noisy observations and imperfect models. Thus, the larger rounding errors incurred
from reducing precision may be within the tolerance of the system. Lower precision
arithmetic is cheaper, and so by reducing precision in ensemble data assimilation, we can
redistribute computational resources towards, for example, a larger ensemble size. Because
larger ensembles provide a better estimate of the underlying distribution and are less reliant
on covariance inflation and localisation, lowering precision could actually allow us to
improve the accuracy of weather forecasts.
We will present results on how lowering numerical precision affects the performance of
an ensemble data assimilation system, consisting of the Lorenz ’96 toy atmospheric model
and the ensemble square root filter. We run the system at half precision (using an
emulation tool), and compare the results with simulations at single and double precision.
We estimate that half precision assimilation with a larger ensemble can reduce
assimilation error by 30%, with respect to double precision assimilation with a
smaller ensemble, for no extra computational cost. This results in around half a
day extra of skillful weather forecasts, if the error-doubling characteristics of the
Lorenz ’96 model are mapped to those of the real atmosphere. Additionally, we
investigate the sensitivity of these results to observational error and assimilation window
length.
Half precision hardware will become available very shortly, with the introduction of
Nvidia’s Pascal GPU architecture and the Intel Knights Mill coprocessor. We hope that the
results presented here will encourage the uptake of this hardware.
References
[1] Peter D. Düben and T. N. Palmer, 2014: Benchmark Tests for Numerical
Weather Forecasts on Inexact Hardware, Mon. Weather Rev., 142, 3809-3829
[2] Peter D. Düben, Hugh McNamara and T. N. Palmer, 2014: The use of imprecise
processing to improve accuracy in weather & climate prediction, J. Comput. Phys.,
271, 2-18 |
|
|
|
|
|