Content uploaded by Sven Kralisch
Author content
All content in this area was uploaded by Sven Kralisch on Mar 26, 2015
Content may be subject to copyright.
International Environmental Modelling and Software Society (iEMSs)
2012 International Congress on Environmental Modelling and Software
Managing Resources of a Limited Planet, Sixth Biennial Meeting, Leipzig, Germany
R. Seppelt, A.A. Voinov, S. Lange, D. Bankamp (Eds.)
http://www.iemss.org/society/index.php/iemss-2012-proceedings
Model representation, parameter
calibration and parallel computing –
the JAMS approach
Sven Kralisch and Christian Fischer
Department of Geoinformatics, Hydrology and Modelling,
Friedrich Schiller University, Jena, Germany
(sven.kralisch@uni-jena.de)
Abstract: To tackle problems of integrated environmental management, flexible
and powerful simulation models are needed in order to analyse the current state of
natural systems or to project their future dynamics under given development sce-
narios. Beyond the mere simulation of physical processes, accompanying tasks like
model calibration or optimization for specific hardware platforms are usually re-
quested here. The Jena Adaptable Modelling System (JAMS) is an open-source
software platform that has been especially designed to address the demands of a
process-based environmental model development and various aspects of model
application. This paper gives an overview of JAMS and its underlying concepts and
shows how its explicit representation of model structure and modelling entities can
support parameter calibration and parallel computing.
Keywords: Environmental modelling; modelling frameworks; model calibration;
parallel computing.
1 INTRODUCTION
Software frameworks and accompanying standards that allow for an easy imple-
mentation and linking of integrated environmental models have gained increasing
attention during the last decade, both from model developers and users. The sys-
tems that have emerged range from pure interface solutions (i.e. focusing more on
coupling already existing models) to simulation frameworks that also cover the cre-
ation of problem-tailored simulation components along given requirements. Re-
views and comparisons of currently available environmental modelling frameworks
and their underlying techniques can be found in Argent (2005), Rizzoli et al. (2008)
and Jagers (2010).
The Jena Adaptable Modelling System (JAMS)
1
is a simulation framework devel-
oped with a thematic priority on hydrological processes (Kralisch and Krause, 2006;
Kralisch et al., 2007). Its focus is not on the coupling of existing environmental
models but on the creation of problem-tailored models from well-defined process
simulation components. These components simulate e.g. interception, potential
evapotranspiration or soil temperature with conceptual or physically based algo-
rithms. Making use of the Java Native Access (JNA) library
2
, JAMS components
may even wrap existing functionality offered via native shared libraries that were
compiled e.g. from C/C++ or Fortran code. With regard to the spatio-temporal do-
main, JAMS aims to simulate environmental processes at discrete points in time
and/or space. Such systems, often referred to as timed event system, are widely-
used by many distributed hydrological models applied in current practice.
1
http://jams.uni-jena.de
2
https://github.com/twall/jna
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
JAMS is a JAVA-based framework, developed under the GNU Lesser General Pub-
lic License. Depending on its application purpose it can be used in different execu-
tion environments (e.g. desktop and server based). In addition to functions for the
creation and application of models, JAMS offers various software tools that cover
frequently requested tasks in the environmental modelling context (e.g. for model
calibration, parameter analysis and result visualization).
The next section will give an overview of the JAMS architecture and its main con-
cepts. Section 3 is dedicated to describe how these support parameter calibration
and parallel processing of JAMS models.
2 JAMS CONCEPTS
2.1 System architecture
The JAMS framework is structured into three main sections, i.e. (i) the core library,
(ii) the runtime system, and (iii) the base component repository (Figure 1). The core
library contains the API definition and different global support functions, e.g. for
loading models or for model data I/O during simulation. The application of the mod-
els is managed by the runtime system, which is responsible for loading and execut-
ing JAMS models and provides additional services for event logging or profiling.
The base component repository
offers standard functions often
used in environmental simula-
tion models, e.g. providing geo-
spatial processing capabilities.
The core library provides inter-
faces and data types both for
the creation of modelling com-
ponents and the runtime sys-
tem. The latter can create and
execute a JAMS model by ac-
cessing the component reposi-
tory and a model definition,
which is defined by a XML doc-
ument.
2.2 Model building blocks
The main building block of any JAMS model is named component. A component is
a JAVA class that implements some given interface defining different methods (init,
run and cleanup). They have to be executed at according runtime stages that every
JAMS model iterates through. Communication with the framework and other com-
ponents is handled by arbitrary public attributes that fulfill two conditions: (i) they are
of a valid JAMS data type and (ii) they are marked by special JAVA annotations, i.e.
syntactic meta-information defining their I/O type (read, write), default values, phys-
ical unit and boundaries (if numeric), and their purpose by means of a short text.
This information is used both by the runtime system to setup and interlink attributes
and by the graphical user interface (GUI) that provides support during model de-
sign.
JAMS components can serve a variety of different purposes. The most important
one is the simulation of environmental processes. An example is the calculation of
potential evapotranspiration (PET), taking wind speed, temperature, humidity, ra-
diation and elevation as input and calculating PET as output, e.g. according to after
Penman-Montheith. Other components typically found in JAMS models implement
data I/O or statistical analyses. They are applied e.g. for reading basic data like
land use and soil type attributes or for aggregating process simulation results over
spatial model entities in a region of interest. The source code complexity of a JAMS
process simulation component can range from less than 50 logical executable lines
of code (SLOC-L) (e.g. for calculating radiation input) up to more than 500 SLOC-L
(e.g. for more complex snow simulation algorithms).
Figure 1. JAMS framework elements
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
A special type of components is a GUI component. In addition to standard compo-
nents they feature a graphical panel that serves as a container for arbitrary GUI
elements. Added to a JAMS model, GUI components can be used to create graph-
ical output during model runtime, e.g. for showing simulation results by utilizing
chart and mapping libraries.
Components in JAMS are stateless objects, meaning that all processing infor-
mation has to be provided via attributes at each invocation of the component’s run
method. Therefore, all state information (e.g. spatial attribute values) must be
stored outside of the component.
Figure 2. Context examples
Although generally possible, temporal or spatial iteration is not meant to be done
inside of components. Instead, this is subject of the second type of model building
blocks, called contexts. A JAMS context is a special, compound-type component
that can nest other components and contexts, named children. It serves three pur-
poses: (i) it controls the execution of its children; (ii) it controls the data exchange
between its children, and (iii) serves as storage for its children’s states. Depending
on the purpose of a given context, it might invoke its children multiple times over a
number of iterations (Figure 2 left), only once if some predefined condition is satis-
fied (Figure 2 center) or only once in a sequence (Figure 2 right). The invocation of
its children is controlled in a context’s run method, leaving full control over the way
they are executed in the hands of the context
developer. Due to this fact, contexts can
address a wide variety of tasks, covering for
example the application of different sub-
models depending on input data and user
requirements or the calibration of models. As
contexts are specialized components they
can be nested in each other, allowing the
creation of complex component hierarchies
and execution control structures.
Spatial and temporal iteration is provided by
the spatial context and the temporal context,
iterating over a set of spatial model entities
or time steps respectively and invoking their
children for each of them. A typical applica-
tion of both contexts is shown in Figure 3,
with a spatial context nested inside of a tem-
poral context. Using this setup, conceptual
models that simulate environmental pro-
cesses at discrete points in space-time can
easily be represented in JAMS.
2.3 Component data exchange
Data exchange between components is managed by their surrounding contexts.
For this purpose, each context features at least one state object which allows the
storage of arbitrary JAMS data. These data are stored in data slots that can be
dynamically added and removed as needed. By connecting their attributes to the
Figure 3. Nesting of temporal
and spatial contexts
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
same data slots, children can easily exchange data during model runtime
(Figure 4).
Sometimes it is necessary that a context can store more than one state, as it is the
case with spatial contexts. In this case, each spatial model entity is reflected by a
separate state, with its spatial attributes (e.g. elevation or slope) stored in data
slots. While iterating over its spatial model entities, the spatial context can restore
an entity’s state prior the invocation of its children and store it afterwards. This way,
a context can provide access to different sets of attributes and thus maintain data
exchange between its children over varying spatial model entities (Figure 4).
Figure 4. Data exchange between components using state objects
The same approach is applied for other context types where it is necessary to store
and restore the state of their children (i.e. their input and output data) during execu-
tion. If there is no need to retain the state, the context uses only one state object,
overwriting the data in its data slots in the case of repeated children invocation.
2.3 Runtime behaviour
A JAMS model is a special context called model context, which is a sequence-type
context as shown in Figure 2 (right). Within this outermost context, other compo-
nents and contexts can be arranged and nested as needed. The execution of the
model is started by iterating through the init, run and cleanup stages of the model
context. This in turn will start the
invocation of its children accord-
ing to a generalized activity
scheme as shown in Figure 5. In
the context’s init stage, its own
init method is invoked first before
it will iterate over its children and
invoke their init methods accord-
ingly. During the run stage, the
context will perform the following
activity sequence depending on
its behavior (e.g. in an iterated,
conditional or sequential fash-
ion): (i) restore the current state,
(ii) iterate over all children and
invoke their run stage, and (iii)
save the current state. After the
run stage has been finished, the
context enters its cleanup stage,
invoking the cleanup stages of
its children first and its own at
the end. This means that in con-
trast to other frameworks, e.g. as
based on the Discrete Event
System Specification (DEVS)
(Zeigler, 2000), the invocation of
Figure 5. Context runtime activities
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
JAMS components is not autonomously triggered by external events, but explicitly
by the surrounding context.
Using this approach, an entire JAMS model (i.e. model context) can easily be re-
used and combined with or nested in other components. A typical use case is mod-
el analysis and calibration, where a model context is added as a child to another
context which allows the automated sampling of parameters of the internal model.
3 MODEL OPTIMIZATION
3.1 Use case J2000
Applying the pattern outlined in Figure 3, environmental models that simulate spa-
tially distributed processes over time can be created easily. An example is the
J2000 hydrological model (Krause, 2002) which has been successfully applied in a
large number of catchment studies covering the lower meso- to lower macro-scale
(Krause & Flügel 2005; Krause et al. 2006). J2000 simulates the water balance of
hydrological catchments based on their spatial decomposition into hydrological
response units (HRUs) (Flügel, 1996) for daily and sub-daily time steps. At each
simulated time step, different runoff components are calculated for each HRU, fol-
lowed by a spatial routing of the lateral runoff components from HRU to HRU or
HRU to stream segment respectively.
Figure 6 shows the generic layout of J2000 in JAMS. The tasks executed during a
simulation with J2000 can be sketched as follows:
1. The model is initialized by reading spatial input data, i.e. a set of HRUs includ-
ing information about their land-use, soil and geological parameters and a set
of stream entities as a basis for streamflow simulation. In addition, the sets of
HRUs and stream segments are analyzed regarding their flow topology and
ordered in such a way that an element is processed only after its contributors
have been processed.
2. At each time step, the needed time series input data are read and steps 3 to 5
are executed.
3. For each HRU, a set
of local input data is
calculated (e.g. by
spatial interpolation
of climate data) and
various hydrological
processes are simu-
lated (e.g. ET, inter-
ception and infiltra-
tion). Lateral runoff
components are
routed to HRUs and
stream segments.
4. For each stream
segment, stream-
flow is simulated.
5. Simulated data are
aggregated and
output.
3.2 Model parallelization
Spatially distributed models often have a high time and space complexity, i.e. de-
pending on the number of spatial model entities and process detail they demand for
large amounts of process memory and computation time. While memory consump-
tion has become less constraining due to the availability of 64-bit operating systems
and continuously dropping hardware costs, the runtime performance of environ-
mental simulation models remains a crucial point. Thanks to multi-core architec-
tures and cloud computing environments, the reason is not so much lacking CPU
Figure 6. J2000 model layout in JAMS
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
power but model algorithms that are badly suited for concurrent processing. In gen-
eral, two requirements must be addressed here which are often closely related:
1. The software platform used must offer support for the concurrent execution of
processing routines.
2. The model’s representation, i.e. its data and algorithms, must allow for an in-
dependent, parallel processing.
For modern software the first requirement is barely a problem as multithreading is a
common paradigm supported by many current programming languages. Further-
more, software interfaces like MPI
3
offer standardized and well-tested support for
distributed execution of simulation routines. The second requirement can become
much more challenging as data and algorithms might induce strong interdependen-
cies between different parts of the model. This problem can often be observed in
distributed environmental models where energy and substances are exchanged
between spatial model entities.
Due to its explicit
representation of
model structure and
model entities, JAMS
offers options to ad-
dress both require-
ments. The concur-
rent execution of
processing routines
is achieved by isolat-
ing sub-models that
can be processed
independently from
each other. A typical
use case for this
scenario is the cali-
bration of simulation
models. This can be
done by embedding the model to be calibrated in a special JAMS context that con-
trols the calibration process (Figure 7). The OPTAS optimization assistant (Fischer
et al., 2009) uses this approach to setup parameter calibration procedures for any
given JAMS model based on a variety of optimization methods and objective func-
tions in a semi-automated way. In the setup shown in Figure 7, the model under
calibration (right) can be executed concurrently for different samples of calibration
parameters as interdependencies are not a problem here. The calibration context
can be provided by the user as every other JAMS component, allowing calibration
methods to be customized to high performance computing environments like com-
puter grids if needed. Applying this approach, the GridGain
4
platform was success-
fully integrated and tested in OPTAS (Fischer et al., 2009).
Looking at an ever increasing amount of available environmental information and
the associated growing complexity of simulation models, parallel processing has
become interesting on a sub-model level, too. Taking into account that multi-core
processing is available in virtually every computer nowadays, this step seems even
more compelling. Models that simulate environmental processes in a spatially dis-
tributed way seem to be a good candidate for parallel computing, but as pointed out
before, attention must be paid to interrelationships between their spatial model enti-
ties.
Applying this approach to JAMS and the J2000 model, only small changes have to
be applied to the model layout, leaving the existing process components un-
touched. In a first step, the set of HRUs is partitioned into n HRU subsets, where n
is the number of concurrent simulation processes. The challenge here is to find
subsets that (i) can be processed independently and (ii) have a similar size. For
hydrological models as J2000, the first requirement can be met by making sure that
HRUs from the same sub-catchment are not distributed into different subsets.
3
http://www.mcs.anl.gov/research/projects/mpi
4
http://www.gridgain.com
Figure 7. Model calibration in JAMS with calibration
context (blue) and J2000 example model (grey)
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
While the second requirement is
not mandatory, a better balanced
partitioning leads to a higher
speedup. Although strongly de-
pending on the characteristics of
the catchment (e.g. number of
HRUs/sub-catchments) and the
number of concurrent processes,
experience has shown that a satis-
fying partitioning can be achieved in
most cases. In a second step, n
copies of the HRU context (cf. Fig-
ure 6) together with all its children
are created and placed within a
special context that replaces the
HRU context (Figure 8). This con-
currency context makes sure that
each of the copies is provided with
one of the previously created HRU
subsets. The conversion of the
model is done by JAMS in a fully automated way, leaving only the implementation
of the concurrency context and the HRU partitioning component in the responsibility
of the user.
The presented approach was tested on a standard workstation computer with a
current 4-core CPU. In order to consider the impact of the number of HRUs on the
achieved speedup, two J2000 models for different catchments were tested. The
first model was represented by 614 HRUs, the second model was about ten times
larger with 6,242 HRUs. A time period of 4 years was simulated with both models,
using simple JAVA multi-threading for concurrent processing.
Figure 9 shows the speedups that were achieved for both models using different
numbers of CPU cores. Speedup is defined as the ratio between compute time
using one core and compute time using multiple cores, i.e. a higher speedup
means less compute time. The red dotted line marks the ideal speedup where the
use of n cores results in 1/n compute time. The data clearly show that J2000 is well
suited for in-model parallel computing, even though only parts of the model were
processed concurrently (i.e. all components within the spatial context). Regarding
the fact that nearly all physical
processes are simulated here
which sums up to about 85% of
the overall compute time (135s for
the larger model), this result is not
that much surprising.
With growing number of cores, the
slopes of the speedup graphs
decline which indicates that an
increase of CPU cores will pay off
only up to a certain point where
the additional computational effort
for parallel computing fully com-
pensates the associated gain. As
indicated in the graph, this point is
reached earlier when using less
HRUs which is explained with the
worse ratio of fixed vs. paralleliza-
ble computational effort.
4 SUMMARY & CONCLUSION
After introducing the main concepts of creating environmental models with JAMS,
special attention was given to the adaptation of models to the requirements raised
by model calibration and parallel computing. Due to the explicit representation of
their structure and spatial modelling entities, JAMS allows to perform this task with-
Figure 9. Speedup of J2000 models
Figure 8. JAMS context for concurrent
processing of HRUs
S. Kralisch & C. Fischer / Model representation, parameter calibration and parallel computing…
out touching their processing components. After looking at parameter calibration
methods, a generalized approach for model-internal concurrent processing was
discussed. This approach uses standard JAMS components only, while existing
process simulation components can be used without modification. The method was
tested with two hydrological models featuring different numbers of modelling enti-
ties and proved to be more effective on the larger model. Ongoing work is focusing
on fine-tuning the partitioning of spatial model entities and on the automated de-
ployment and parallel processing in cloud computing environments.
ACKNOWLEDGEMENTS
The authors acknowledge the support of the German Ministry of Education and
Research which has funded the JAMS development as part of the InnoProfile pro-
gram (grant number 03IP514).
REFERENCES
Argent, R.M., 2005. A case study of environmental modelling and simulation using
transplantable components. Environmental Modelling & Software 20, 1514–
1523.
Fischer, C., Kralisch, S., Krause, P., Fink, M., Flügel, W.-A., 2009. Calibration of
hydrological model parameters with the JAMS framework, in: Anderssen,
R.S., Braddock, R.D., Newham, L.T.H. (Eds.), Proceedings of the 18th
World IMACS Congress and MODSIM09 International Congress on Model-
ling and Simulation. Cairns, Australia, pp. 866–872.
Flügel, W.-A., 1996. Hydrological response units (HRU’s) as modelling entities for
hydrological river basin simulation and their methodological Potenzial for
modelling complex environmental process systems - Results from the Sieg
catchment. Die Erde 127, 43–62.
Jagers, H.R.A. (Bert), 2010. Linking Data, Models and Tools: An Overview, in:
Swayne, D.A., Yang, W., Voinov, A.A., Rizzoli, A., Filatova, T. (Eds.), Pro-
ceedings of the iEMSs Firth Biennial Meeting: International Congress on
Environmental Modelling and Software (iEMSs 2010). International Envi-
ronmental Modelling and Software Society, Ottawa, Canada.
Kralisch, S., Krause, P., 2006. JAMS - A Framework for Natural Resource Model
Development and Application, in: Voinov, A., Jakeman, A., Rizzoli, A.E.
(Eds.), Proceedings of the iEMSs Third Biannual Meeting. Burlington, USA.
Kralisch, S., Krause, P., Fink, M., Fischer, C., Flügel, W.-A., 2007. Component
based environmental modelling using the JAMS framework, in: Kulasiri, D.,
Oxley, L. (Eds.), Proceedings of the MODSIM 2007 International Congress
on Modelling and Simulation. Christchurch, New Zealand.
Krause, P., 2002. Quantifying the impact of land use changes on the water balance
of large catchments using the J2000 model. Physics and Chemistry of the
Earth 27, 663–673.
Krause, P., Bende-Michl, U., Bäse, F., Fink, M., Flügel, W.-A., Pfennig, B., 2006.
Multiscale Investigations in a Mesoscale Catchment – Hydrological Model-
ling in the Gera Catchment. Advances in Geosciences 9, 53–61.
Krause, P., Flügel, W.-A., 2005. Integrated research on the hydrological process
dynamics from the Wilde Gera catchment in Germany, in: Headwater Con-
trol VI: Hydrology, Ecology and Water Resources in Headwaters. Present-
ed at the IAHS Conference, Bergen.
Rizzoli, A.E., Leavesley, G., Ascough II, J.C., Argent, R.M., Athanasiadis, I.N.,
Brilhante, V., Claeys, F.H.A., David, O., Donatelli, M., Gijsbers, P., Havlik,
D., Kassahun, A., Krause, P., Quinn, N.W.T., Scholten, H., Sojda, R.S., Vil-
la, F., 2008. Chapter Seven Integrated Modelling Frameworks for Environ-
mental Assessment and Decision Support, in: Environmental Modelling,
Software and Decision Support. Elsevier, pp. 101–118.
Zeigler, B., 2000. Theory of modeling and simulation : integrating discrete event and
continuous complex dynamic systems, 2nd ed. ed. Academic Press, San
Diego.
