Reference Manual & User's Guide

Reference Manual & User's Guide
Deliverable D1.3:
NEMO Tangent & Adjoint Models
(NemoTam)
Reference Manual & User’s Guide
Arthur Vidard, Franck Vigilant
INRIA/LJK, Grenoble
Charles Deltel, Rachid Benshila
LOCEAN-IPSL, Paris
VODA: ANR-08-COSI-016
VODA: ANR-08-COSI-016
2
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
Contents
1
2
N EMOTAM R EFERENCE M ANUAL
1.1 Introduction . . . . . . . . . . . . . . . . . . . .
1.1.1 Abstract . . . . . . . . . . . . . . . . . .
1.1.2 Introduction . . . . . . . . . . . . . . . .
1.2 Tangent and adjoint models (TAM) basics . . . .
1.2.1 TAM, but what for ? . . . . . . . . . . .
1.2.2 Tangent and Adjoint coding principle . .
1.2.3 Potential issues . . . . . . . . . . . . . .
1.3 Direct model changes . . . . . . . . . . . . . . .
1.3.1 Storing the non-linear trajectory . . . . .
1.3.2 Modified routines in the direct code . . .
1.3.3 Modified declarations in the direct code .
1.4 Tangent linear model approximations . . . . . . .
1.4.1 Active variables degraded to passive . . .
1.4.2 Simplifications . . . . . . . . . . . . . .
1.5 Hand coding status . . . . . . . . . . . . . . . .
1.6 Available options in NEMOTAM . . . . . . . . .
1.7 The generic validation interface . . . . . . . . .
1.7.1 Introduction . . . . . . . . . . . . . . . .
1.7.2 Adjoint test . . . . . . . . . . . . . . . .
1.7.3 Tangent-Linear validation . . . . . . . .
1.7.4 Tangent-Linear Hypothesis (HLT) validity
1.8 MPI parallelization . . . . . . . . . . . . . . . .
1.8.1 General features . . . . . . . . . . . . .
1.8.2 Adjoint of MPI ALLGATHER . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
N EMOTAM U SER ’ S G UIDE
2.1 Source code structuration . . . . . . . . . . . . . . . . . . . . . .
2.2 Getting started tutorial . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Installing the code using the NEMOVAR GIT repository .
2.2.2 Installing the code using the NEMOTAM SVN repository
2.2.3 Which executable for which test? . . . . . . . . . . . . .
2.3 Demonstrators . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
5
5
7
7
7
8
8
8
10
10
10
10
11
11
11
16
16
16
17
19
20
20
21
.
.
.
.
.
.
23
23
24
24
25
26
26
VODA: ANR-08-COSI-016
2.4
2.5
4
2.3.1 The ORCA2 global configuration . . . . . . . . . . . . . . . . . . . . .
2.3.2 The GYRE regional configuration . . . . . . . . . . . . . . . . . . . . .
2.3.3 The POMME configuration (toy model for handling open boundaries) . .
How-to update NEMOTAM with your contribution? . . . . . . . . . . . . . . . .
2.4.1 Coding rules for TAM . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Workflow for introducing additional developments . . . . . . . . . . . .
Frequently asked questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1 I can’t use GDB after compiling with FCM . . . . . . . . . . . . . . . .
2.5.2 I have many commas in the namelist generated by piano . . . . . . . . .
2.5.3 I have negative values in my cost function terms . . . . . . . . . . . . .
2.5.4 My run stops with ’PROBLEM WITH THE MATRIX OR WITH THE
PRECONDITIONER’ on stderr . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
26
32
33
34
34
35
36
36
36
36
. 36
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
Chapter 1
N EMOTAM R EFERENCE M ANUAL
1.1
Introduction
This document still contains empty parts. This is done on purpose since it will be the basis for the
next deliverable (D1.3 NEMOTAM Manual). We will fill them in as we proceed.
1.1.1
Abstract
The development of the tangent linear and adjoint models (TAM in the following) of the dynamical
core of the NEMO ocean engine (NEMOTAM) is a key objective of the VODA project. TAM are
widely used for variational assimilation applications, but they are also powerful tools for the analysis of physical processes, since they can be used for sensitivity analysis, parameter identification
and for the computation of characteristic vectors (singular vectors, Liapunov vectors, etc.).
In the framework of VODA, a work package has been set-up in order to develop a comprehensive NEMOTAM package, and to define an effective long-term development strategy for ensuring
synchronisation of NEMOTAM with future NEMO releases. This is a heavy task, but it is worth the
effort since NEMOTAM will benefit all NEMO users for the wide range of applications described
above.
Ideally, this strategy should be defined to allow NEMOTAM to adapt to future NEMO developments as quickly and as efficiently as possible, so that new releases of NEMOTAM can be made
soon after new releases of NEMO. This will require careful coordination between the main development teams of NEMO, NEMOTAM and possibly NEMOVAR (INRIA, NEMO Team, CERFACS, ECMWF).
1.1.2
Introduction
The NEMO ocean engine (?) was previously known as the OPA model (?). It used to have a
TAM (called OPATAM), fully hand-coded and maintained mainly by A. Weaver. OPATAM was
initially developed for a Pacific ocean configuration, and targeted at variational data assimilation
applications in the framework of OPAVAR (??). OPATAM/OPAVAR were extended to other regional basins (Mediterranean sea (?), North Atlantic 1/3◦ (?), South Atlantic 1◦ ), to the global
5
VODA: ANR-08-COSI-016
ocean (ORCA 2◦ (?)), and were used for methodological studies such as control of the 3D model
error (?), control of the surface forcing and open boundary conditions (??). OPATAM was also
used for sensitivity studies (?), singular vectors (??), etc.
For several reasons, mainly because of lack of workforce, OPATAM, OPAVAR and related
developments were not included in the standard release of OPA. As a consequence, synchronisation
of OPATAM with OPA’s releases could not be achieved on a regular basis, and all developments
were on individual branches, without feedback to the OPATAM/OPAVAR system. The pool of
potential users was reduced consequently. It is important not to repeat this error in the future, so
as to ensure that NEMOTAM become a widely used community tool.
A NEMOTAM working group was initiated in the framework of a CNRS/INSU/LEFE ASSIM2006 project, to investigate the feasibility of using TAPENADE (?), an AD tool, to speed up the
writing of TAM for OPA9, the dynamical ocean component of NEMO. The goal of this working
group was twofold. The first goal was to identify the strengths and weaknesses of TAPENADE
by applying it directly to the NEMO source code, with as little human intervention as possible, in
order to build a NEMOTAM prototype. The second goal was to define, based on the experience
deriving the prototype, a strategy for developing a general-purpose NEMOTAM that both respects
the NEMO code style/structure and gives acceptable computer performance (in terms of both CPU
and memory requirements) for realistic ocean configurations. Providing feedback to the TAPENADE developers was an important aspect of this work to help improve future versions of the
TAPENADE software.
The results from this feasibility study demonstrated that TAPENADE was able to produce
tangent-linear and adjoint models for NEMO, albeit for a fixed and somewhat simplified configuration (ORCA 2◦ with all non-differentiable options switched off (?)). Even for this simplified
configuration, however, substantial human intervention and additional work was required to obtain a useable product from the raw TAPENADE-generated code. Three main drawbacks with
TAPENADE were identified for this application. First, the memory management and CPU performance of the raw code were rather poor. Second, the current version of TAPENADE generates
single-processor code only and cannot handle directives from the C-PreProcessor (CPP keys),
which are widespread in NEMO. Third, the technique of binomial checkpointing that is used in
TAPENADE to handle nonlinearities (see (?)) is not compatible, at least in its present implementation, with the incremental algorithm of NEMOVAR, which employs separate executables and
(possibly) different resolutions for the outer and inner loops. Improved memory management and
extensions to support MPP and CPP keys are planned in future versions of TAPENADE so the
first two deficiencies are not fundamental. The third deficiency, however, is more problematic and
it is likely that the trajectory management for nonlinearities in NEMOTAM will be done differently from TAPENADE, possibly along the lines of the simpler strategy implemented in OPAVAR.
The modifications required to make TAPENADE or whatever other AD tool, compatible with the
multi-incremental approach are really substantial and cannot be done in a short or medium term.
Moreover the numerical performances of the TAPENADE generated TAM do not allow yet their
use for ’big’ configurations and for operational applications.
From that experience it has been decided to go toward the hand-coding approach, the use of
AD tool being left aside for the time being. We may reconsider this in a medium or long term
though.
6
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
1.2
1.2.1
Tangent and adjoint models (TAM) basics
TAM, but what for ?
The development of tangent and adjoint models is an important step in addressing sensitivity analysis and variational data assimilation problems in Oceanography. Sensitivity analysis is the study
of how model output varies with changes in model inputs. The sensitivity information given by
the adjoint model is used directly to gain an understanding of the physical processes. In data assimilation, one considers a cost function which is a measure of the model-data misfit. The adjoint
sensitivities are used to build the gradient for descent algorithms. Similarly the tangent model is
used in the context of the incremental algorithms to linearize the cost function around a background
control.
1.2.2
Tangent and Adjoint coding principle
The original program P , whatever its size and run time, computes a function
F : X ∈ Rm → Y ∈ Rn
which is the composition of the elementary functions computed by each run-time instruction. In
other words if P executes a sequence of elementary statements Ik , k ∈ [1, .., p], then P actually
evaluates
F = fp ◦ fp−1 ◦ ... ◦ f1
where each fk is the function implemented bu Ik . Therefore one can apply the chain rule of
derivative calculus to get the Jacobian matrix F 0 , i.e. the partial derivatives of each component of
Y with respect to each component of X. Calling X0 = X and Xk = fk (Xk−1 ) the successive
values of all intermediate variables, i.e. the successive states of the memory throughout execution
of P , we get
0
F 0 (X) = fp0 (Xp−1 ) × fp−1
(Xp−2 ) × ... × f10 (X0 )
(1.1)
The derivatives fk0 of each elementary instruction are easily built, and must be inserted in the
differentiated program so that each of them has the values Xk−1 directly available for use. This
process yields analytic derivatives, that are exact up to numerical accuracy. In practice, two sorts
of derivatives are of particular importance in scientific computing: the tangent (or directional)
derivatives, and the adjoint (or reverse) derivatives. The tangent derivative is the product
dY = F 0 (X) × dX
of the full Jacobian times a direction dX in the input space. From equation (1.1), we find
0
dY = F 0 (X) × dX = fp0 (Xp−1 ) × fp−1
(Xp−2 ) × ... × f10 (X0 ) × dX
(1.2)
which is most cheaply executed from right to left because matrix × vector products are much
cheaper than matrix × matrix products. This is also the most convenient execution order because
it uses the intermediate values Xk in the same order as the program P builds them. On the other
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
7
VODA: ANR-08-COSI-016
hand the adjoint derivative is the product Xad = F 0? (X) × Yad of the transposed Jacobian times a
weight vector Yad in the output space. The resulting Xad is the gradient of the dot product (Y.Yad ).
From equation (1), we find
0?
Xad = F 0? (X) × Yad = f10? (X0 ) × ... × fp−1
(Xp−2 ) × fp0? (Xp−1 ) × Yad
(1.3)
which is also most cheaply executed from right to left. However, this uses the intermediate values
Xk in the inverse of their building order in P .
For specific details about practical adjoint coding refer to (?).
1.2.3
Potential issues
• the major approximations : non differentiability issues, simplification of the direct model
before linearization, validity of the tangent linear hypothesis over some time window, etc.
• the validation interface : classical tests for checking the correctness of tangent and adjoint
models
• numerical problems
• numerical cost issues
1.3
Direct model changes
Some modifications and additions are compulsory to integrate the Tangent and Adjoint Module to
NEMO. They mainly consist in:
• changing subroutines or parameters declaration from private to public for:
– initialization
– re-use
• adding a subroutine to save the non-linear trajectory
• adding parameters
• updating namelist
1.3.1
Storing the non-linear trajectory
To evaluate the tangent and/or the adjoint at a time t, some actual parameters of the direct model
are needed at the corresponding time t. To fulfill this purpose a subroutine, tamtrj wri is added to
write the trajectory. It is controlled with one logical parameter: ln trjwri. The trajectory is saved
upon a given frequency ”nittrjfrq” only if the logical parameter is set to TRUE.
New subroutine tam trj wri
A new subdirectory TAM is created into OPA SRC named TAM. The file is tamtrj.F90. The
purpose of this routine is to write on disk the model state trajectory for use with 4DVar. It is called
in the opa model routine for the initializiation, and in the stp routine in order to perform the
actual saving of the trajectory. The currently stored fields are listed in Table 1.1.
8
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Parameters
emp
emps
un
vn
tn
sn
ta
sa
tb
sb
avmu
avmv
aeiu
aeiv
aeiw
uslp
vslp
wslpi
wslpj
avs
strdmp
hmlp
Associated ccp keys
lk traldf eiv
lk traldf eiv
lk traldf eiv
key ldfslp
key ldfslp
key ldfslp
key ldfslp
key zdfddm
key tradmp
key tradmp
Table 1.1: Fields stored in the reference trajectory
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
9
VODA: ANR-08-COSI-016
Specific management of the initial state
For the initial state of the model, we call the tam trj wri in the ’step’ routine just before the tracer
block. We, then, have access to some parameters (tke and sbc for instance) as they are not yet
computed in opa (outside ’step’).
Linear interpolation
For efficiency matters, the whole trajectory is of course not saved. Only a predefined subset of
the time steps is saved. The frequency on which the trajectory is saved is controlled by namelist
parameter nittrjfrq. The intermediate steps will be estimated in TAM by the mean of a linear
interpolation.
Namelist
Parameters related to the model state saving is controlled by the namelist namtam with:
• ln trjwri: Logical switch for writing out state trajectory
• nittrjfrq: saving frequency of the trajectory
1.3.2
Modified routines in the direct code
• opa.F90:
– include initilialization of the module that handles the trajectory saving (USE tamtrj)
• step.F90:
– include calls for saving the trajectory (CALL tam trj wri)
– add temporary variables (zta, zsa) to save ta/sa (this is a temporary fix because these
variables are currently used as workspace in dynamic block)
1.3.3
Modified declarations in the direct code
The following table describes the change from PRIVATE to PUBLIC of some routines and parameters.
1.4
1.4.1
Tangent linear model approximations
Active variables degraded to passive
Some active variables are treated as passive (mainly due to non-differentiability characteristics).
• avmu, avmv: the vertical eddy viscosity. The turbulent closure scheme zdftke induces
some non-differentiabilities for there variables. They are used in routine dynzdf_imp_tam.
F90.
10
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Modules
dynadv
daymod
lib mpp
par oce
zdftke
eosbn2
tradmp
Routines
dyn adv ctl
day mth
Parameters
mpi comm opa
jpni, jpnj, jpnij
tke rst
neos init
cofdis, dtacof, dtacof zoom
traqsr
strdmp, ttrdmp, resto
nksr, gdsr
Table 1.2: List of modified declarations in the direct code NEMO
• uslp, vslp: the slope of neutral surface (slope of isopycnal surfaces referenced locally) are
treated as passive in order to not recomputed them at each step. They are used in routine
traldf_iso_tam.F90.
• wslpi/j: vertical mixing coefficient due to lateral mixing. They are used in routine trazdf_
imp_tam.F90.
1.4.2
Simplifications
• removal of the upstream part of the second order centered scheme for traadv_cen2_tam.
F90. It introduces some non-differentiabilities. Some complementary test must be held to
fully implement the second order scheme
• removal of zonal mean lateral diffusive heat and salt transport (pht ldf and pst ldf )
1.5
Hand coding status
The current status for the hand coding part of NEMOTAM / NEMOVAR is gathered in Tables 1.3
and 1.4.
1.6
Available options in NEMOTAM
This sections contains information on current available configuration for NEMOTAM. To have
an overview on available routines, please see the TAM development progress table. We use the
following method: The physical choices are set through cpp keys and namelist parameters. We use
below a similar table used for NEMO. See tables 1.5 and 1.6.
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
11
VODA: ANR-08-COSI-016
Modules called
by step tam
Module(s) called
by previous one
Options
Remarks
Ocean dynamics (DYN)
cla div tam.F90
cla tam.F90
divcur tam.F90
dynadv tam.F90
dynkeg
dynzad
dynadv
dynadv
tam.F90
tam.F90
cen2 tam.F90
ubs tam.F90
not in reference
not in reference
dynhpg tam.F90
hpg zco hpg zps
hpg sco hpg hel
hpg wdj, hpg djc, hpg rot
dynldf tam.F90
dynldf
dynldf
dynldf
dynldf
lap tam.F90
bilap tam.F90
iso tam.F90
bilapg tam.F90
not in reference
not in reference
dynnxt tam.F90
lk vvl
key obc, key bdy
key agrif
non-linear
not in reference
not in reference
dynspg tam.F90
dynspg flt tam.F90
dynspg ts tam.F90
dynspg exp tam.F90
not in reference
not in reference
dynvor tam.F90
dyn
dyn
dyn
dyn
vor
vor
vor
vor
ens
een
ene
mix
not in reference
not in reference
dynzdf tam.F90
dynzdf exp tam.F90
dynzdf imp tam.F90
param. avmu and avmv
as passive variables (if not:
non-differentiability issue,
see zdftke formulation)
(should be treated as
active var. when
key zdf noncst is activated
dynmem tam.F90
wzvmod tam.F90
Table 1.3: Hand coding status of NEMOTAM, Dynamics part
12
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Modules called
by step tam
Module(s) called
Options
by previous one
Ocean tracers (TRA)
Remarks
traadv tam.F90
no upstream scheme
(equivalent to zcofi=0)
not differenciable
not differentiable
not differentiable
not differenciable
not differentiable
not differentiable
traadv cen2 tam.F90
traadv
traadv
traadv
traadv
traadv
traadv
tvd tam.F90
muscl tam.F90
muscl2 tam.F90
ubs tam.F90
qck tam.F90
eiv tam.F90
trabbc tam.F90
hlmp is saved in direct
model to handle uslp, vslp
tradmp tam.F90
traldf tam.F90
key ldf ano
uslp, vslp treated as passive
not in reference
not in reference
pht ldf, pst ldf not include
key obc, key bdy
key agrif
not in reference
not in reference
traldf bilap tam.F90
traldf bilapg tam.F90
traldf lap tam.F90
tranxt tam.F90
traqsr tam.F90
trasbc tam.F90
trazdf tam.F90
lk vvl
non-linear
wslpi, wslpj treated as passive
trazdf imp tam.F90
lk vvl
trazdf exp tam.F90
lk vvl
Surface boundary conditions (SBC)
sbcmod tam.F90
sbc
sbc
sbc
sbc
sbc
sbc
gyre tam.F90
flx tam.F90
ssr tam.F90
fwb tam.F90
clo tam.F90
ana tam.F90
Table 1.4: Hand coding status of NEMOTAM, Tracers and Surface Boundary Condition part
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
13
VODA: ANR-08-COSI-016
Used for
Horizontal
Diffusion/Viscosity
Functionalities
CPP keys
Lateral diffusion along
isopycinal slopes
Associated namelist logical
key ldfslp
Horizontal
Tracer diffusion
Tracer lateral diffusion eddy
induced velocity parametrization
Tracer lateral laplacian
Iso-neutral application direction
for operator
Tracer lateral diffusion
1d=f(depth), 2d=f(lat,ion) or
3d=f(lat,ion, depth)
Horizontal
Dynamics Viscosity
key traldf eiv
ln traldf lap
ln traldf iso
key traldf c2d
nam dynldf
Dynamic Laplacian operator
lateral viscosity
Dynamic biLaplacian operator
lateral viscosity
Geopotential application direction
viscosity
Dynamic lateral viscosity
1d, 2d or 3d
ln dynldf lap
ln dynldf bilap
ln dynldf lap
key dynldf 3cd
Vertical viscosity
Diffusivity
Vertical diffusion based on TKE
closure scheme
Vertical double diffusion mixing
for salinity
key zdftke
namtke
key zdfddm
namddm
Vertical convective
processes
enhanced vertical diffusion
parametrization
ln zdfevd
namsbc,
key forced daily,
key tau daily
Surface forcing
fields
Initial state
and damping
Internal Newtonian tracers
(T,S) damping
Use a 3D monthly salinity
data (Levitus)
Use a monthly temperature
data (Levitus)
Use a monthly sst
data (Levitus)
key tradmp
key dtasal
key dtatem
key dtasst
Sea surface temperature
damping
14
Sea surface salinity
damping
Deliverable
namdtp
D1.3: NEMOTAM Reference
namsbc,
namsbc ssr,
ln ssr
namsbc,
namsbc ssr,
Manual
ln ssr & User’s Guide
Table 1.5: Available option in NEMOTAM, part1
VODA: ANR-08-COSI-016
Used for
On-line diagnostics
Functionalities
Eddy induced velocity
(Gent-Williams parametrization)
Thermocline, 20◦ C-28◦ C
isotherm depths
Enable observation operator
CPP keys
Associated namelist logical
key diaeiv
key diahth
key diaobs
High Performances
computing
Enable MPP computing
with MPI
reproductibility
key mpp mpi
key mpi send
key mpi send
Vertical coordinates
nam zgr
ln zco
ln zps
Full steps - Z
Partial steps - PS
Pressure gradient
treatment
Filtered free surface algorithm
key dynspg flt
Vorticity scheme
nam dynvor
ln dynvor ens
ln dynvor een
nam traadv
ln traadv cen2
enstrophy conserving
energy and enstrophy conserving
Advection schemes
2nd order centered
Configuration
Idealized Double Gyre
configuration with a flat bottom
Global ORCA configuration
at 2deg resolution
Open boundaries POMME
configuration at 0.25◦ resolution
key gyre
key orca r2
key pomme r025
Other CPP keys
Geothermal boundary condition flux
Diffusive Bottom Boundary Layer
key trabbl dif
nambbc
nambbl
Table 1.6: Available option in NEMOTAM, part2
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
15
VODA: ANR-08-COSI-016
1.7
1.7.1
The generic validation interface
Introduction
The validation interface aims at showing the consistency of the TAM with respect to the direct
code. To fulfill this goal we demonstrate that:
1. the adjoint code is the adjoint of its tangent code
2. the tangent linear code is the tangent linear of the direct code
with defined criteria.
In general, TAM modules are subdivided into four parts:
1.
2.
3.
4.
a tangent-linear routine
an adjoint routine
an adjoint test routine
a tangent-linear routine
Item 3 and 4 are optional. The strategy is to perform adjoint and tangent-linear tests for all critical
sub-modules and the main module step tam for time-step integration.
1.7.2
Adjoint test
By definition, A∗ (or tA in a finite-dimension case) is the adjoint operator of a linear operator A
defined on a Hilbert space if it has the following property:
(Ax, y) = (x, tAy)
(1.4)
In practice, we apply the comparison to random perturbations δx in the model-space and δy ∗ in the
adjoint model-space. We have then:
(Aδx, δy ∗ ) = (δx, tAδy ∗ )
(1.5)
Adjoint test requirements
• A saved trajectory of the direct model
• An additionnal namelist (namtst for adjoint test, namtst tlm for tangent test
Unitary test versus Global test
Unitary test involves one or a limited set of routines. The pertubation vectors δX and δY are the
input/output variables identified as actives variables. There is no time evolution. Global test involves time evolution (loop on step tam) and δX and δY are input/output variables identified as
the control vector (or variables of main interest).
16
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Test Pass Criteria
We note (if exists) the relative error Err. The pass test criteria is based on the comparison of Err
with the epsilon machine εmachine = eps.
Err =
(Aδx, δy ∗ ) − (δx, tAδy ∗ )
(Aδx, δy ∗ )
(1.6)
The adjoint test results output is the file adjoint test.output xxx with the following label:
Test status is OK
test status is WARNING
Test status is FAILED
Err ≤ Eps
Eps < Err ≤ 100 × Eps
Err > 100 × Eps
Table 1.7: Adjoint test labels definition
Run the adjoint test
• set the directory where the direct trajectory is stored
• in the namelist namtst
–
–
–
–
1.7.3
set ln
set ln
set ln
set ln
tst
tst
tst
tst
nemotam to .true. (switch for M adjoint and tangent tests)
cpd tam to .true. (switch for adjoint tests of the sub-components)
stp tam to .true. (switch for adjoint tests of main components step tam)
tan cpd and ln tst tan to .false. (switch for tangent tests)
Tangent-Linear validation
The tangent validation test checks that the Tangent Linear model L is a first-order approximation
of the direct model M. Then, we have the following second-order Taylor expansion:
M(X + pδX) = M(X) + pL(δX) + ε
(1.7)
When the factor p tends to zero:
1. Np = M(X + pδX) − M(X) tends to Lp = pL(δX) , first-order validation
2. ε behaves as O(p2 ) , second-order validation
this means that for a given p0 , for p smaller than p0 the error of Tangent Linear with respect to the
model is decreasing as p2 .
Note that the second test (b) is not relevant if M is linear.
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
17
VODA: ANR-08-COSI-016
Test
a
b
Expression
N orm(Np , .)
N orm(Lp , .)
N orm(Np − Lp , .)
p2
behaviour
1
Cst
Table 1.8: Tangent Linear tests. First-order check (a) and second-order check (b)
Build the tangent test
The tangent test is embedded into nemotam. To build it the user must add the compilation key
key tst tlm in the CPP key set. Otherwise, a error message will raise during the run.
Run the tangent test
To run the tangent test the code must
• set the directory where the direct trajectory is stored
• in the namelist namtst
–
–
–
–
set ln
set ln
set ln
set ln
tst to .true.
tst nemotam to .true. (switch for M adjoint and tangent tests)
tst tan cpd or ln tst tan to .true. (switch for tangent tests)
tst cpd tam and ln tst stp tam to .false. (switch for adjoint tests)
• in the namelist namtst tlm
– set tlm bch to .true. (branching: ’0’ for M(X), ’1’ for M(X + pδX) and ’2’ for LpdX )
– set curr loop to .true. (current iteration)
– set h ratio to .true. (value of the ratio factor p applied to δX)
!----------------------------------------------------------------------!
namtst_tlm
tangent test parameters
!----------------------------------------------------------------------!
! cur_loop = current loop iteration
! h_ratio = current h_ratio
! tlm_bch = flag for branching
!
= 0 :M(X)
, direct without perturbation
!
= 1 :M(X+h_ratio.dX), direct with perturbation
!
= 2 :L(h_ratio.dX) , linear
&namtst_tlm
tlm_bch = 1
cur_loop = 0
h_ratio = 10
/
18
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Tangent test ouputs
The output is saved with the name tan diag.output 0000. The header give information on the computed parameters. For iteration n, the output gives the follwing data:
routine name, Nn = M(X+pn δX)−M(X), En = Nn /L(pn δX), Ern = (Nn −L(pn δX))/L(pn δX),
L(pn δX), Nn − L(pn δX), (En − 1)/pn , Ern /pn .
To have an overview of the behaviour of the tangent iteration can be made by the user changing
the factor ratio pn . The factor takes usually a geometric progression as: pn = p0 r−n , with initial
value p0 and common ratio r.
1.7.4
Tangent-Linear Hypothesis (HLT) validity
The innovation vector validity relies on the validity of the Tangent Linear. Before any 4DVar assimilation we must insure that the tangent model is valid within our assimilation time-window.
Otherwise the optimisation may fail as the cost function gradient will be degraded or even wrong.
We define a maximum assimilation time-window Tamax . This time-window will be different for
each configuration. According to each application the user its own criteria to define to define Tamax
to fulfill its needs. The size of Tamax may vary from 1 to 3 months (???).
To estimate Tamax we study the behaviour of y1 = Mt0 →t (x0 +δx) with respect to y2 = Mt0 →t (x0 )+
Lt0 →t (δx).
The Tangent Linear will be valid as long as the difference y1 − y2 is ”small”. As discussed in
(?) definition criteria of Tamax will rely on threshold on the RMS of y1 − y2 and the coefficient
correlation ρy1 y2 . A correlation coefficient ρy1 y2 above 90% is considered as very good.
RM S(y1 − y2 ) =
v
n
uX
u
u
(y1i − y2i )2
t
i
n
, n: number of ocean points
(1.8)
n
X
(y1i − y¯1 ) · (y2i − y¯1 )
s n
ρ y1 y2 = s n i
X
X
(y1i − y¯1 )2 ·
(y2i − y¯2 )2
i
, n: number of ocean points
(1.9)
i
The initial condition as well as the increment δx are critical to define Tamax . The user should
work with a typical increment δx of its configuration.
HLT test tool
NEMOTAM offers the possibility to analize the Tangent-Linear Hypothesis through a dedicate executable model hlt.exe. The increments δx is either given by the user either computed from two
restarts file x0 and x1 (then δx = x0 − x1 ). To compute y1 and y2 , model hlt.exe should be run
three times: two for the direct model trajectories (Mt0 →t (x0 + δx) and Mt0 →t (x0 )) and one for
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
19
VODA: ANR-08-COSI-016
the tangent trajectory (Lt0 →t (δx)). It is driven through the namelists namhlt and tl tamtrj.
!----------------------------------------------------------------------!
namhlt
tangent-linear hypothesis test parameters
! nstg
flag for to compute model with or without perturbation
!
0: M(X)
!
1: M(X+dX)
!
2: L(dX)
! ln_hltt
Logical switch for applying temperature increments
! ln_hlts
Logical switch for applying salinity increments
! ln_hltuv
Logical switch for applying horizontal velocity increments
! ln_hltssh Logical switch for applying sea surface height increments
! ln_incdx
Logical swith to compute dx (F) or reading from file (T)
! ln_hnorm
Logical swith to normalization (T) or not (F) using rhstdX
! rhstdt
Upper bound of normalization factor for temperature
! rhstds
Upper bound of normalization factor for salinity
! rhstduv
Upper bound of normalization factor for velocity
! rhstdssh
Upper bound of normalization factor for sea surface height
!----------------------------------------------------------------------&namhlt
nstg = 2
ln_hltt
= .TRUE.
ln_hlts
= .TRUE.
ln_hltuv = .TRUE.
ln_hltssh = .FALSE.
c_hltinc = ""
ln_incdx = .FALSE.
ln_hnorm = .TRUE.
rhstdt
= 1.
rhstds
= 0.1
rhstduv = 0.01
rhstdssh = 0.01
/
!----------------------------------------------------------------------!
namtl_trj
tangent-linear trajectory frequency output
! ln_trjwri_tan Logical switch rajectory output
! nittrjfrq_tan frequency output for linear tangent
!----------------------------------------------------------------------&namtl_trj
ln_trjwri_tan = .TRUE.
nittrjfrq_tan = 96
/
1.8
1.8.1
MPI parallelization
General features
We remind some basics for coding the corresponding adjoint of the tangent linear part:
20
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Operation
communication
communication
communication
arithmetic
inpout/Output
inpout/Output
Tangent-Linear
send
receive
gather
scatter
write
read
Adjoint
receive
send
scatter
gather
read
write
Table 1.9: ORCA2 Results of Adjoint test fo sub-routines
There is no tangent of lbc lnk. The direct routine is used. For the adjoint part we use the coding
rule described in table 1.9
1.8.2
Adjoint of MPI ALLGATHER
From revision 3.2 of NEMO a new treatment of north fold horizontal bondary condition is performed using lbcnfd.F90 routine. This induces a MPI ALLGATHER usage in libmpp.F90 module
(mpp lbc north 3d and mpp lbc north 2d).
MPI ALLGATHER gathers data from all tasks and distribute it to all.
Figure 1.1: MPI ALLGATHER task description (ref: http://rc.usf.edu/tutorials/classes
/tutorial/mpi/chapter8.html
If we call zloc Pi the data from task (i) and the zglo the result of MPI ALLGATHER and
zloc adj Pi , zglo adj their adjoint equivalent.
The adjoint counter-part of
MPI_ALLGATHER(zglo, zloc)
is
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
21
VODA: ANR-08-COSI-016
MPI_SUM(zglo_adj)
zloc__Pi = zglo_adj(:,i)
As short demonstration, we consider a two-task case P1 and P2 .
MPI ALLGATHER(zglo, zloc) is equivalent to:
for P1 :
(1) send and receive local data:
send(zloc1_P1, P1, P2)
--->
recv(zloc2_P1, P2, P1)
--->
zloc1_P2 = zloc1_P1
zloc2_P1 = zloc2_P2
( zlocK_Pj : local data K of j task)
B- zglo operation
zglo_P1(:,1) = zloc1_P1
zglo_P1(:,2) = zloc2_P1
we then write the adjoint as follows:
zloc2_P1*
zglo_P1*(:,2)
zloc1_P1*
zglo_P1*(:,1)
=
=
=
=
zloc2_P1* + zglo_P1*(:,2)
0
zloc1_P1* + zglo_P1*(:,1)
0
zglo_tot*
= zglo_P1* + zglo_P2*
(i.e mpp_sum_nfd)
(! because zloc_P1* is not null here, thus: zloc1_P1* = zloc_P1* + zloc1_P2*)
zloc1_P1*
zloc2_P1*
22
= zglo_tot*(:,1)
= zglo_tot*(:,2)
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
Chapter 2
N EMOTAM U SER ’ S G UIDE
2.1
Source code structuration
The project will work with three repositories (see Fig. 2.1b,c,d):
• a NEMOBASE repository, with an official NEMO released version currently the tag nemo v3,
available at :
https://forge.ipsl.jussieu.fr/nemo/browser/tags/nemo_v3
• a NEMOVAR repository, with all developments related to the nemovar project, including
new and modified files for NEMO, this is the standard version of NEMOVAR (currently tag
N3.0-A2.0)
• a VODA repository including new development for NEMOTAM and NEMOVAR within the
VODA project
(a) GIT tree
(b) NEMOBASE
(c) NEMOVAR
Figure 2.1: Code organization
23
(d) VODA
VODA: ANR-08-COSI-016
The building process will look for both repositories, if any redundancy is found, a VODA
files will overwrite the corresponding NEMOVAR file, and a NEMOVAR file will overwrite the
corresponding NEMOBASE file in the building folder.
The NEMOBASE repository (see Fig. 2.1b) is the standard NEMO folder. The NEMOVAR
repository (see Fig. 2.1c) is subdivided into 6 components: NEMO, NEMOQC, NEMOTAM,
NEMOASSIM, UTIL and TOOL. The NEMOTAM folder is a mimic of the NEMO folder (meaning we have an OPATAM SRC folder subdivided into components as ASM, DOM, DYN, LDF,
TRA, ...). The VODA folder (see Fig. 2.1d) mimic the NEMOVAR folder, with additional contributions. This organization gives a clear snapshot of what is being developped and lets some room
for further development.
2.2
2.2.1
Getting started tutorial
Installing the code using the NEMOVAR GIT repository
Checking out the code from the GIT repository
You need to ask the VODA project team to have developer access to the VODA project on the
gforge.inria.fr server and access to the NEMO - IPSL. You can then follow the steps described
below.
Create first a new directory, where the source code will be installed.
For instance,
cd $VODA_GIT (=$HOME/VODA-PROJECT/VODA_GIT e.g.)
---> Install NEMOBASE
go to https://forge.ipsl.jussieu.fr/nemo/browser/tags/nemo\_v3 (authentication
required)
and click on ’zip archive’ at the bottom of the page.
put this archive in $VODA_GIT, unzip it and rename the obtained directory
NEMOBASE (this name is mandatory)
---> Install NEMOVAR
Download the following bundles, located on the gforge server under the "files"
tab:
NEMOVAR: N3.0-A2.0 => nemovar_N3.0-A2.0.bundle
and put them in the $VODA_GIT/BUNDLES directory.
In $VODA_GIT, you now have:
tree BUNDLES/
BUNDLES/
|-- NEMOBASE.bundle.20090902
‘-- nemovar_N3.0-A2.0.bundle
Clone these bundles :
git clone BUNDLES/NEMOBASE.bundle.20090902 NEMOBASE/
git clone BUNDLES/nemovar_N3.0-A2.0.bundle NEMOVAR/
24
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
---> Install VODA
git clone git+ssh://[email protected]//gitroot/voda/voda.git VODA/
--> Final structure of the $VODA_GIT directory
tree -d -L 2 $VODA_GIT
VODA_GIT
|-- BUNDLES
|-- NEMOBASE
|
|-- AGRIF
|
|-- CONFIG
|
|-- CVSROOT
|
|-- DOC
|
|-- IOIPSL
|
|-- NEMO
|
|-- NVTK
|
‘-- UTIL
|-- NEMOVAR
|
|-- NEMO
|
|-- NEMOASSIM
|
|-- NEMOQC
|
|-- NEMOTAM
|
|-- TOOLS
|
‘-- UTIL
‘-- VODA
|-- NEMO
|-- NEMOASSIM
|-- NEMOTAM
‘-- UTIL
Building the code (in default mode : ORCA2 validation interface)
NEMOTAM makes use of the fcm based NEMOVAR compiling environment, a simple call to
VODA/UTIL/build/fcmvmake.ksh with the usual arguments and the nemoall keyword should do
the trick.
> fcmvmake.ksh -c $COMPILE -t $WORKDIR -B nemoall
the above requires that you create a VODA/UTIL/build/fcmconfig/bld/$COMPILE/nemo.cfg describing your compilation environment (examples are given in the same directory).
How to run the code is described in section 2.3.
2.2.2
Installing the code using the NEMOTAM SVN repository
Checking out the code from the SVN repository
NEMOTAM can be extracted using svn like NEMO. People should first to register on NEMO website www.nemo-ocean.eu. Then NEMOTAM is available through modipsl framework, so modipsl
has to be extracted first.
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
25
VODA: ANR-08-COSI-016
• Extract modipsl:
svn co http://forge.ipsl.jussieu.fr/igcmg/svn/modipsl/trunk modipsl
• Extract NEMOTAM:
cd modipsl/util
./model NEMOTAM
Building the code (in default mode : ORCA2 validation interface)
Choose the configuration to build. You can edit the script fait config tam to add a new one.
cd modipsl/modeles/UTIL
./fait_config_tam ORCA2_TAM
You can build the direct model in the same way:
cd modipsl/modeles/UTIL
./fait_config ORCA2_LIM
To create the Makefiles, edit first modipsl/util/AA make.gdef to find or add your compiler, then:
cd modipsl/util
./ins_make
To compile:
cd modipsl/modeles/NEMOTAM/WORK
gmake
2.2.3
Which executable for which test?
To run a bunch of tests to verify your configuration you need to compile several executable according to your test target. The graph 2.2 give an overview of the relationship between the executable
and the available tests.
Running the code
2.3
2.3.1
Demonstrators
The ORCA2 global configuration
ORCA is the generic name given to global ocean configurations. Its specificity lies on the horizontal curvilinear mesh used to overcome the North Pole singularity found for geographical meshes.
ORCA is based on a 2 degrees Mercator mesh, (i.e variation of meridian scale factor as cosinus of
the latitude). In the northern hemisphere the mesh has two poles so that the ratio of anisotropy is
nearly one everywhere. The vertical domain spreads from the surface to a depth of 5000m. There
are 31 levels, with 10 levels in the to 100m. The time step depends on the resolution. It is 1h360
for ORCA2 so that there is 15 time steps in one day.
26
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
EXE file
model tam.exe
nemovar inner.exe
Adjoint test
OPATAM
ln tst=T,
ln tst nemotam=T
ln tst cpd tam, ln tst stp tam
Tangent test
OPATAM
ln tst=ln tst tan=T
Adjoint test
BKG, BAL,
OBS
ln tst=T,
ln tst nemotam=T
Gradient Test
ln tst cpd tan
ln tst obsadj, ln tst bkgadj
ln tst=T,
ln tst nemotam=T
ln tst grd
Single
Obs. exp.
model hlt.exe
Linear Tangent
Hypothesis Test
ln sgl=T
namhlt
nstg=1,2,3
Figure 2.2: Releationships between executable files (first column) and available tests (second column). The thrid column focuses on the main namelist’s parameters (’namtst’ for adjoint, gradient,
tangent tests, ’namobs’ for single obs.experiement and ’namhlt’ for linear tangent hypothesis)
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
27
VODA: ANR-08-COSI-016
How to compile the ORCA2 configuration ?
You must first follow the steps described in 2.2.1. The compilation process described below then
relies upon the FCM (Flexible Configuration Manager) toolbox, which is included in VODA tree,
see e.g. the following directory :
ls $VODA_GIT/VODA/UTIL/fcm/bin
Then, you may proceed in the following manner:
• Get sample input data, needed for the model to run (http://ljk.imag.fr/membres/Arthur
.Vidard/docs/nemov3_testdata.tar.gz) :
• Customize the first configuration file nemo.cfg, which is located according to your platform
architecture (mac intel here) :
vi $VODA_GIT/VODA/UTIL/build/fcmconfig/bld/mac_intel/nemo.cfg
• Customize the second configuration file, with CPP keys: cppkeys.ORCA2 Z31.cfg. A sample file is ready to be used, which you can modify:
vi $VODA_GIT/VODA/UTIL/build/fcmconfig/nemo/cppkeys.ORCA2_Z31.cfg
• You are now ready to compile the code. You need to create a ”scratch” directory (for example
SCRATCH ORCA2 below), where compilation and execution will be performed:
cd $VODA_GIT/VODA/UTIL/build
./fcmvmake.ksh -t $SCRATCH_ORCA2 -c mac_intel -B NEMO
-G ORCA2_Z31
./fcmvmake.ksh -t $SCRATCH_ORCA2 -c mac_intel -B NEMOTAM -G ORCA2_Z31
How to run the ORCA2 configuration ?
A sample minimal script is given, just as an example : The direct model NEMO
$VODA_GIT/VODA/UTIL/scripts/runnemo.ksh
NEMO
-g ORCA2_Z31
-t $SCRATCH_ORCA2 -A mac_intel -v
The adjoint model NEMO tests
$VODA_GIT/VODA/UTIL/scripts/runnemo.ksh
TRAJ
-v NEMOTAM -g ORCA2_Z31
-t $SCRATCH_ORCA2 -A mac_intel -j $
where $TRAJ is the reference trajectory directory. This script absolutely needs to be customized
with respect to your working environment. Please have a look at this script before proceeding
further. You will first need to create a reference trajectory by running only the direct model NEMO,
and you then may want to use NEMOTAM to
• launch the generic TAM test interface,
• investigate the time span over which the tangent linear hypothethis remains valid,
• etc.
Testing the tangent model and the tangent linear hypothesis
As discussed previously, the aims of the test is to illustrate the correctness of the tangent and how
we have dealt with the non-linearity (if it exits) of the direct model.
28
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Figure 2.3: Tangent test type: Four kinds of test results. Top left (a), linear routine with a residual
following the error machine. Top right (b), non-linear routine with a residual decreasing as porder of 2. Bottom left (c), linear routine with a residual decreasing as p-order of 1. Bottom
right (d), non-linear routine with a residual decreasing as p-order of 1. (Legend: Solid line is
Np = M (x + pδx) − M (x), circle symbol Lp = L(pδx), triangle symbol Np − Lp )
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
29
VODA: ANR-08-COSI-016
Discussion
We have divided the results into four categories (a), (b), (c) and (d) for analysis (see table 2.1 ).
Category (a), illustrated by solsor tan in figure 2.3, gathers tangent-linear routines behaving as
expected w.r.t to their direct routine (also linear) counter-part: the computed residual is close to the
Lp
error machine (the difference
is around 10−15 . Category (b), illustrated by eos insitu tan
Np − Lp
in figure 2.3, gathers tangent-linear routines behaving as expected w.r.t to their direct routine (not
linear) counter-part: the computed residual is equivalent to a O(p2 ). Categories (c) and (d) , illustrated by dynspg flt tan and tra adv cen2 tan in figure 2.3, gather suspicious cases we need to
validate (or invalidate) through the choice made for the tangent-linear. Case (c) is for direct linear
routine and (d) direct non-linear routine.
Routine (L)
solsor tan
dynadv tan
dynhpg tan
traldf lap tan
trazdf imp tan
tra zpshde tan
trasbc tan
bn2 tan
eos 1pt tan
eos insitu2d tan
eos insp tan
eos insitu tan
dynspg flt tan
traadv cen2 flt tan
Category class
a
b
a
b
b or c ?
b
b
b
b
b
b
b
c
d
Table 2.1: ORCA2 Results of Tangent tests for sub-routines
The case of traadv cen2 flt tan is fully understood as an approximation was done on the numerical scheme (the tangent linear does not include the upstream component of the numerical
scheme). Below a certain threshold the upstream part (which is a O(p) in Np ) is no longer small
compared to the second order scheme and thus, Lp .
Testing the adjoint model
The adjoint test for the ORCA2 configuration has proven satisfactory for the unitary routine tests
(see table 2.2 ) and the global test (see tables 2.3 and 2.4. For the latter, the integration time tested
varied from 1 to 15 days.
30
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Routine (L)
sol sor adj
sbcfwb adj 1 1
sbcfwb adj 2 1
sbcfwb adj 1 2
sbcfwb adj 2 2
sbcfwb adj 1 3
sbcfwb adj 2 3
sbc ssr adj
sbc gyre adj 1
sbc gyre adj 2
sbc ssm adj
bn2 adj
div cla adj
div cur adj T1
div cur adj T2
dyn adv adj
dynvor adj ens
dynvor adj een
dynldf adj lap
dynldf adj blp
dyn zdf adj
dyn spg flt T1
dyn spg flt T2
dyn hpg adj
dyn nxt adj
wzv adj
tra sbc adj
tra qsr adj
tra qsr adj
tra dmp adj T1
tra dmp adj T2
tra dmp adj T3
tra adv cen2
tra cla adj Gi
tra cla adj BM
tra ldf iso ad
trazdf exp adj
trazdfimpadjT1
trazdfimpadjT2
trazdfimpadjT3
tra nxt adj
eos adj insitu
eos adj pot
eos adj 2d
eos adj 1pt
zps hde adj
istate tst
(Lδx)| · W dy
.525406844734452E+15
.711617836790924E+12
.711586863123878E+12
.721405227628168E+12
.721331456200554E+12
.711617836790924E+12
.711586863123878E+12
.289793337319649E+21
.575805928604001E+19
.575805928604001E+19
.585782123944269E+15
.763641537078732E+10
.134110702040363E+17
.307633707985862E+07
.321541201767843E+07
.252818496373574E+17
.252818508381207E+17
.252818506922870E+17
.311731217229964E+17
.252575821916828E+17
.252556994319522E+17
.790035995408866E+24
.315978990829171E+25
.253953510430215E+17
.758385105599429E+17
.258438017479678E+13
.517847966550809E+15
1 .134738569058199E+19
2 .134738569058199E+19
.152739703065678E+19
.152739703065678E+19
.152739703065678E+19
.152558770733990E+19
.487513953864278E+16
.158030397951644E+15
.152906929763072E+19
.235296147886893E+37
.208237542411001E+25
.208237542411001E+25
.208237542411001E+25
.405901725366058E+19
.405930021470975E+11
.241698506116296E+17
.249470161251995E+09
.108202320860649E+10
.417898655654515E+18
.140086193684597E+19
δdx| · L| W δy
.525406844734452E+15
.711617836790924E+12
.711586863123878E+12
.721405227628168E+12
.721331456200554E+12
.711617836790924E+12
.711586863123878E+12
.289793337319649E+21
.575805928604001E+19
.575805928604001E+19
.585782123944269E+15
.763641537078733E+10
.134110702040363E+17
.307633707985862E+07
.321541201767843E+07
.252818496373574E+17
.252818508381207E+17
.252818506922870E+17
.311731217229965E+17
.252575821916828E+17
.252556994319522E+17
.790035995408866E+24
.315978990829171E+25
.253953510430214E+17
.758385105599429E+17
.258438017479678E+13
.517847966550810E+15
.134738569058200E+19
.134738569058200E+19
.152739703065678E+19
.152739703065678E+19
.152739703065678E+19
.152558770733990E+19
.487513953864278E+16
.158030397951644E+15
.152906929763072E+19
.235296147886893E+37
.208237542411001E+25
.208237542411001E+25
.208237542411001E+25
.405901725366058E+19
.405930021470975E+11
.241698506116297E+17
.249470161251995E+09
.108202320860649E+10
.417898655654515E+18
.140086193684597E+19
Rel. Err.
.0E+00
.7E-15
.0E+00
.0E+00
.0E+00
.7E-15
.0E+00
.2E-15
.0E+00
.0E+00
.0E+00
.5E-15
.1E-15
.0E+00
.6E-15
.2E-15
.0E+00
.0E+00
.3E-15
.2E-15
.0E+00
.0E+00
.3E-15
.3E-15
.0E+00
.0E+00
.5E-15
.1E-14
.1E-14
.0E+00
.0E+00
.0E+00
.1E-14
.2E-15
.6E-15
.0E+00
.6E-15
.5E-15
.5E-15
.5E-15
.9E-15
.0E+00
.5E-15
.4E-15
.2E-15
.2E-15
.7E-15
Mach. eps.
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
Status
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
ok
Table 2.2: ORCA2 Results of the Adjoint tests for the main sub-routines
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
31
VODA: ANR-08-COSI-016
Routine (L)
step adj U
step adj V
step adj T
step adj S
step adj SSH
step adj
(Lδx)| · W dy
.161275281471349E+18
.318648865958325E+18
.132120109179127E+19
.281362832449564E+18
.509250348807475E+10
.206971044341451E+19
δdx| · L| W δy
.161275281471349E+18
.318648865958325E+18
.132120109179127E+19
.281362832449564E+18
.509250348807474E+10
.206971044341451E+19
Rel. Err.
.6E-15
.4E-15
.0E+00
.0E+00
.2E-14
.1E-15
Mach. eps.
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
Status
ok
ok
ok
ok
ok
ok
Table 2.3: ORCA2 Results of the Adjoint tests for the main routine step tam for a run-window of
one day. From line one to four, each control parameter is perturbed individually. Column six, all
control parameter are perturbed.
Routine (L)
step adj U
step adj V
step adj T
step adj S
step adj SSH
step adj
(Lδx)| · W dy
.102643067009108E+18
.167827302823834E+18
.893406215846414E+18
.220114453479612E+18
.265003707314821E+10
.137645166426475E+19
δdx| · L| W δy
.102643067009108E+18
.167827302823834E+18
.893406215846413E+18
.220114453479612E+18
.265003707314821E+10
.137645166426475E+19
Rel. Err.
.2E-15
.4E-15
.7E-15
.4E-15
.4E-15
.6E-15
Mach. eps.
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
.2E-14
Status
ok
ok
ok
ok
ok
ok
Table 2.4: ORCA2 Results of the Adjoint tests for the main routine step tam for a run-window of
five days. From line one to four, each control parameter is perturbed individually. Column six, all
control parameter are perturbed.
2.3.2
The GYRE regional configuration
GYRE is an idealized configuration representing double gyres in the Northern hemisphere. It is
β-plane with a regular grid spacing at 1 degree horizontal resolution. There is 31 verticles levels.
The configuration is forced with analytical heat, freswater and wind-stress fields.
How to compile the GYRE configuration ?
Similar to ORCA2 configuration replacing ORCA2 Z31 by GYRE Z31.
How to run the GYRE configuration ?
Similar to ORCA2 configuration replacing ORCA2 Z31 by GYRE Z31.
Tangent-Linear Hypothesis (HLT) validity
To illustrate the tangent linear hypothesis validity test We use an initial state x0 ten years from spinup and an increment δx computed from an previous assimiilation with (|δxT | ≤ 1 , |δxS | ≤ 0.1,
|δxU | ≤ 0.0022 and |δxV | ≤ 0.0019) . see figure 2.4
here: place more stats on δx (mean, min, max, std) ?
here: add comments for the graph
32
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
Figure 2.4: Tangent linear hypothsesis validty for GYRE Z31: plot in function of time-window
in days over 1 year. Top graphs are RMS of T, S, U and V variables in S.I units. Bottom graphs
are related coefficient correlation.
2.3.3
The POMME configuration (toy model for handling open boundaries)
Description of the configuration
NEMO is distributed with some reference configurations allowing the users to set up a first application, and the developers to validate their developments. POMME is a small square box configuration (30 × 40 × 46 grid points), extracted from NEMO DRAKKAR / ORCA025. All boundaries
are open ocean boundaries (including the corners). This configuration was set up to facilitate developments and tests for regional configurations (however, it is not yet an official NEMO reference
configuration). It should be soon included in the NVTK validation toolkit.
How to compile the POMME configuration ?
You must first follow the steps described in 2.2.1. The compilation process described below then
relies upon the FCM (Flexible Configuration Manager) toolbox, which is included in VODA tree,
see e.g. the following directory :
ls $VODA_GIT/VODA/UTIL/fcm/bin
Then, you may proceed in the following manner:
• Get sample input data, needed for the model to run:
wget http://www.locean-ipsl.upmc.fr/˜cdlod/data/data_POMME_Z46.tar.gz
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
33
VODA: ANR-08-COSI-016
• Customize the first configuration file nemo.cfg, which is located according to your platform
architecture (mac intel here) :
vi $VODA_GIT/VODA/UTIL/build/fcmconfig/bld/mac_intel/nemo.cfg
• Customize the second configuration file, with CPP keys: cppkeys.POMME Z46.cfg. A
sample file is ready to be used, which you can modify:
vi $VODA_GIT/VODA/UTIL/build/fcmconfig/nemo/cppkeys.POMME_Z46.cfg
• You are now ready to compile the code. You need to create a ”scratch” directory (for example
SCRATCH POMME below), where compilation and execution will be performed:
cd $VODA_GIT/VODA/UTIL/build
./fcmvmake.ksh -t $SCRATCH_POMME -c mac_intel -B NEMO
-G POMME_Z46
./fcmvmake.ksh -t $SCRATCH_POMME -c mac_intel -B NEMOTAM -G POMME_Z46
How to run the POMME configuration ?
A sample minimal script is given, just as an example :
vi $VODA_GIT/VODA/UTIL/scripts/run_pomme.ksh
This script absolutely needs to be customized with respect to your working environment. Please
have a look at this script before proceeding further. You will first need to create a reference trajectory by running only the direct model NEMO, and you then may want to use NEMOTAM to
• launch the generic TAM test interface,
• investigate the time span over which the tangent linear hypothethis remains valid,
• etc.
Testing the tangent model and the tangent linear hypothesis
coming soon
Testing the adjoint model
coming soon
2.4
2.4.1
How-to update NEMOTAM with your contribution?
Coding rules for TAM
NEMOTAM modules should follow the NEMO convention. However additional rules has to be
followed:
• For a given direct module mod.F90 (that contains the subroutine sub), a corresponding module mod tam.F90 has to be created and it should contain the tangent linear subroutine called
sub tan, the adjoint subroutine sub adj and the corresponding testing routine sub adj tst.
34
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
• For a given active direct variable var, the corresponding tangent linear and adjoint variable
should be named var tl and var ad respectively. The underscore is omitted for local active
variable. To summarized, active variables un, pun, zun will give un tl, pun tl and zun tl in
the tangent module
• For historical reasons, some modules do not follow these rules, feel free to update them.
If the TAM module is a direct translation of the corresponding OPATAM routine, a good practice
is to account for it in the first line of the history (i.i stating that the 8.2 version was done by e.g
Huey, Dewey, and Louie)
2.4.2
Workflow for introducing additional developments
coming soon
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
35
VODA: ANR-08-COSI-016
2.5
Frequently asked questions
2.5.1
I can’t use GDB after compiling with FCM
When I run the code with GDB for debugging purposes, the library lib fcm nemotam.a is not
found, and thus gdb does not have access to symbolic informations associated with the modules of
the code : we can’t set breakpoints, we can’t look for the content of variables, etc. This library is a
temporary one generated by FCM during the build process, but it is removed once the compilation
is finished.
This problem only appears on Mac OSX’s computers. For unknown reasons, the linker looses
track when going from fcm’s temporary libraries to the final libraries. The temporary workaround
is to recreate these temporary libraries from the object files (but you will have to redo it each time
you compile):
cd $SCRATCH_POMME/fcmbuild/POMME_Z46_mac_intel.1.1/build/lib/
ar -q lib__fcm__nemotam.a ../obj/*.o
2.5.2
I have many commas in the namelist generated by piano
This happens if your config namelist have lines ending with a comma. This feature is accepted by
FORTRAN, but the PIANO python script parser do not handle this. Removing the commas at the
end of each line will solve the problem.
2.5.3
I have negative values in my cost function terms
See congrad.F90 :
Compute cost for diagnostic purposes only
pf=zf0+0.5*DOT_PRODUCT(...)
It is this term in congrad which appears to be < 0 and is written to the cost file, whereas we would
expect to write the pcost from simvar.F90. Under investigation
2.5.4
My run stops with ’PROBLEM WITH THE MATRIX OR WITH
THE PRECONDITIONER’ on stderr
This problem was reported in the POMME and GYRE configurations.
First, it is worth noting that this message disappears if you set
ln_tst_grad = .TRUE.
which is absolutely not a normal behaviour. Needs further investigation.
The problem is solved by increasing the namelist nmin parameter. For example, change
nmin
=
300
!
minimum of iterations for the SOR solver
=
500
!
minimum of iterations for the SOR solver
into :
nmin
36
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
VODA: ANR-08-COSI-016
in namsol. This behaviour is not normal, it suggests a convergence problem in the tangent barotropic
solver. Needs further investigation.
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
37
VODA: ANR-08-COSI-016
38
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
List of Figures
1.1
MPI ALLGATHER task description (ref: http://rc.usf.edu/tutorials/classes
/tutorial/mpi/chapter8.html . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1
2.2
Code organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Releationships between executable files (first column) and available tests (second
column). The thrid column focuses on the main namelist’s parameters (’namtst’ for
adjoint, gradient, tangent tests, ’namobs’ for single obs.experiement and ’namhlt’
for linear tangent hypothesis) . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tangent test type: Four kinds of test results. Top left (a), linear routine with a residual following the error machine. Top right (b), non-linear routine with a residual
decreasing as p-order of 2. Bottom left (c), linear routine with a residual decreasing as p-order of 1. Bottom right (d), non-linear routine with a residual decreasing
as p-order of 1. (Legend: Solid line is Np = M (x + pδx) − M (x), circle symbol
Lp = L(pδx), triangle symbol Np − Lp ) . . . . . . . . . . . . . . . . . . . . . .
Tangent linear hypothsesis validty for GYRE Z31: plot in function of timewindow in days over 1 year. Top graphs are RMS of T, S, U and V variables
in S.I units. Bottom graphs are related coefficient correlation. . . . . . . . . . .
2.3
2.4
39
. 23
. 27
. 29
. 33
VODA: ANR-08-COSI-016
40
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
List of Tables
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
Fields stored in the reference trajectory . . . . . . . . . . . . . . . . . . . . . . .
List of modified declarations in the direct code NEMO . . . . . . . . . . . . . .
Hand coding status of NEMOTAM, Dynamics part . . . . . . . . . . . . . . . .
Hand coding status of NEMOTAM, Tracers and Surface Boundary Condition part
Available option in NEMOTAM, part1 . . . . . . . . . . . . . . . . . . . . . . .
Available option in NEMOTAM, part2 . . . . . . . . . . . . . . . . . . . . . . .
Adjoint test labels definition . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tangent Linear tests. First-order check (a) and second-order check (b) . . . . . .
ORCA2 Results of Adjoint test fo sub-routines . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
2.1
2.2
2.3
ORCA2 Results of Tangent tests for sub-routines . . . . . . . . . . . . . . . . .
ORCA2 Results of the Adjoint tests for the main sub-routines . . . . . . . . . . .
ORCA2 Results of the Adjoint tests for the main routine step tam for a run-window
of one day. From line one to four, each control parameter is perturbed individually.
Column six, all control parameter are perturbed. . . . . . . . . . . . . . . . . .
ORCA2 Results of the Adjoint tests for the main routine step tam for a run-window
of five days. From line one to four, each control parameter is perturbed individually. Column six, all control parameter are perturbed. . . . . . . . . . . . . . . .
. 30
. 31
2.4
41
9
11
12
13
14
15
17
18
21
. 32
. 32
VODA: ANR-08-COSI-016
APPENDIX : Acronyms
• AD : Automatic differentiation
• AGRIF : Adaptive Grid Refinement In Fortran. Outil de raffinement de maillage et de nesting
d´evelopp´e au sein de MOISE et inclus dans NEMO.
• ESOPA : Equipe Syst`eme OPA, qui g´erait les d´eveloppements de OPA. A e´ t´e remplac´ee par
la NEMO-Team apr`es la mise en place du consortium NEMO
• MITgcm : Mod`ele de circulation g´en´eral pour l’oc´ean et l’atmosph`ere d´evelopp´e au MIT
• NEMO : Nucleus for European Modelling of the Ocean.
• NEMOTAM : TAM pour NEMO
• NEMO-Team : Equipe g´erant les d´eveloppements de NEMO
• NEMOVAR : syst`eme d’assimilation variationnelle de donn´ees pour NEMO
• OPA : Ocean Parall´elis´e. Le mod`ele direct d’oc´ean, composant de NEMO.
• OPATAM : TAM pour OPA 8.x
• OPAVAR : syst`eme d’assimilation variationnelle de donn´es pour OPA 8.x
• ORCA : grille globale pour OPA
• PONGO : Pˆole d’Oc´ean Num´erique GrenOblois, regroupant les e´ quipes MOISE (LJK/INRIA) et MEOM (LEGI)
• ROMS : Regional Ocean Modelling System
• TAF : Transformation of Algorithms in Fortran. G´en´erateur automatique de code tangent et
adjoint utilis´e pour MITgcm.
• TAM : Mod`eles Tangent et Adjoint
• TAPENADE : G´en´erateur automatique de code tangent et adjoint d´evelopp´e par le projet
TROPICS de l’INRIA
42
Deliverable D1.3: NEMOTAM Reference Manual & User’s Guide
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement