1 |
heimbach |
1.1 |
Poster Title: An efficient exact adjoint of the parallel MIT general |
2 |
|
|
circulation model, generated via automatic differentiation |
3 |
|
|
|
4 |
|
|
Authors: Patrick Heimbach(1), Chris Hill(1), Ralf Giering(2) |
5 |
|
|
Institutions of authors (in author order): (1): Department of Earth, |
6 |
|
|
Atmospheric and Planetary Sciences, Massachusetts Institute of |
7 |
|
|
Technology, Cambridge, MA 02139, USA |
8 |
|
|
(2): FastOpt, Martinistr. 21, 20251 Hamburg, Germany |
9 |
|
|
|
10 |
|
|
Abstract: We describe computational aspects of automatic differentiation |
11 |
|
|
applied to global ocean circulation modeling and state estimation. |
12 |
|
|
The task of minimizing a cost function measuring the ocean simulation |
13 |
|
|
vs. observation misfit is achieved through efficient calculation of |
14 |
|
|
the cost gradient w.r.t. a set of controls via the adjoint technique. |
15 |
|
|
The adjoint code of the parallel MIT general circulation |
16 |
|
|
model is generated using TAMC or its successor TAF. |
17 |
|
|
The adjoint can be generated for a variety of configurations, including |
18 |
|
|
different mixing schemes such as KPP and GM, time-varying surface flux |
19 |
|
|
or atmospheric state controls, and open boundary controls. |
20 |
|
|
To achieve a tractable problem in both CPU and memory requirements, in |
21 |
|
|
the light of control flow reversal, the adjoint code relies heavily on |
22 |
|
|
the balancing of storing vs. recomputation via the checkpointing method. |
23 |
|
|
Further savings are achieved by exploiting self-adjointedness of part |
24 |
|
|
of the computation. To retain scalability of domain decomposition |
25 |
|
|
based parallelism, hand-written adjoint routines are provided. |
26 |
|
|
These complement routines of the parallel support package |
27 |
|
|
to perform corresponding operations in reverse mode. |
28 |
|
|
A unique feature of the TAF tool which enables to dump the adjoint |
29 |
|
|
state and restart the adjoint integration is exploited to overcome |
30 |
|
|
batch execution limitations on HPC machines for large-scale ocean and |
31 |
|
|
climate simulations. |
32 |
|
|
The size of a typical adjoint application is illustrated for the global |
33 |
|
|
ocean state estimation problem. Results for a sensitivity study and an |
34 |
|
|
estimation problem are given by way of example. |
35 |
|
|
|