![Example optimization loop involving heterogeneous simulation tasks in its workflow [4]. Workflow description: The perceived comfort is evaluated by a Human Response Model (HRM), provided as a web-service by external partner C. The HRM requires inputs of temperature, air flow, humidity and pressure (ENV) from the Cabin CFD model and noise and vibration (N&V) inputs from the Cabin FEM model. The ENV results of the Cabin CFD model are calculated based on its own operational characteristics in combination with those of the ECS electrical/thermal model from external partner A. The N&V results of the Cabin FEM model are calculated based on its own operational characteristics in combination with those of the power plant FEM model from external partner B. Different settings and configurations for each of the four models (FEM, CFD, electric/thermal) are considered based on the optimization loop algorithm and the available constraints, the results of which are then evaluated by the HRM, leading to selection of new values for calculation and evaluation. The loop continues until the selected criteria have been met.](https://www.researchgate.net/profile/Vassilios-Moussas/publication/280939582/figure/fig1/AS:613888764694528@1523373895007/Example-optimization-loop-involving-heterogeneous-simulation-tasks-in-its-workflow-4_Q320.jpg)
Content uploaded by Vassilios C Moussas
Author content
All content in this area was uploaded by Vassilios C Moussas on Apr 10, 2018
Content may be subject to copyright.
6th International Conference from “Scientific Computing to Computational Engineering”
6th IC-SCCE
Athens, 9-12 July, 2014
© IC-SCCE
AN ALGORITHM FOR WEB-BASED DISTRIBUTED HETEROGENEOUS
SIMULATION WORKFLOW MULTI-OBJECTIVE OPTIMIZATION
Tsahalis J.1, Moussas V.C.1,2 and Tsahalis H.-T.1
1 Paragon S.A.
Karaoli & Dimitriou 13, Galatsi, GR-11146, Athens, Greece
e-mail: jtsahalis@paragon.gr, web page: http://www.paragon.gr
2 School of Technological Applications, Tech. Educ. Inst. (TEI) of Athens,
Ag. Spyridonos Str., Egaleo 12210, Athens, Greece
e-mail: vmouss@teiath.gr, web page: http://users.teiath.gr/vmouss/
Keywords: Optimization methods; scheduling; simulation workflow; evolutionary algorithms; web services;
distributed tools; heterogeneous network;
Abstract. The increased interest on processing large scale & heterogeneous problems in distributed
environments created the need of software tools that would support such complex workflows. Especially,
simulation workflow scheduling has become an important area as it allows users to process large scale
problems in a more flexible way. In most complex simulation workflows the user has to select the optimal use of
local and external resources that will satisfy its requirements under the specific time & cost constraints. In this
work we present a Simulation Workflow Optimization (SWO) algorithm that is based on heuristic optimization
techniques (Genetic Algorithms) and delivers an optimized workflow implementation of an initial plan or
workflow schedule. The aim of SWO is to address the increased complexity encountered when one or more
distributed & heterogeneous processes are involved in a simulation workflow. A heterogeneous simulation
workflow contains several virtual tasks that involve completely different software tools, resources, requirements
and often contradictory objectives. In addition, the distributed environment of large scale problems requires the
software tools to be accessible from anywhere as been local. In order to support remotely the solution of each
specific optimization problem, the SWO algorithm is developed as: a) a web based tool designed to function in a
distributed environment and invoked using web services, and b) a tool that can be specialized per task, domain,
product or application by means of knowledge bases, ontologies and user provided information.
1 INTRODUCTION
1.1 Workflows & Simulation Workflows
A workflow is defined as “a reliably repeatable pattern of activity enabled by a systematic organization of
resources, defined roles and mass, energy and information flows, into a work process that can be documented and
learned. Workflows are always designed to achieve processing intents of some sort, such as physical
transformation, service provision, or information processing.” [1].
Workflow optimization must take into account multiple objectives and constraints of high, medium and low
levels of each workflow. In order to address the workflow optimization efficiently in the current work, the
following two-level approach has been adopted:
• Human workflow: represents the higher (i.e. generalized) level of a workflow. Will typically represent a
schedule based on the defined tasks and their dependencies as well as the associated resources (human,
physical, virtual, etc.). During the optimization process of human workflows, the lower-level costs and
constraints (e.g. accuracy) associated with simulations are generic (i.e. not specific to the case under
examination) and considered as non-variable (hard), e.g. a specific CFD simulation is considered to have
a specific computational cost that is not variable depending on the specific circumstances under which it
is required to run.
• Simulation workflow: represents the lower (i.e. detailed) level of a workflow. In this case, the associated
lower-level costs and constraints associated with simulations are considered as variable, enabling further
optimization for the specific case under examination. This is achieved by utilizing semantic annotations
of each model and simulation task. For the CFD simulation example above, this means that simulation
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
criteria such as accuracy can be varied to investigate for the optimal balance between e.g. acceptable
accuracy, computational cost and simulation duration that fit the draft schedule.
Human workflow optimization is applied first in the overall optimization process to generate a draft optimal
schedule of simulations to be performed. Then the simulation workflow optimization generates optimal
propositions for individual simulation configurations. Once the administrator selects an optimal configuration,
the human workflow optimization is run once again to revise the original schedule to the new simulation-related
costs and constraints, thus providing the optimal application-specific test planning. In this sense a configuration
can be the overall set of human and simulation workflow configurations. In this way, the administrator’s
decisions are ‘supported’ by the available algorithms, but the final decision is still the administrator’s control.
A simple example workflow of weather forecasting is presented in figure 1 below, displaying the sequence of
concatenated steps along with their relevant descriptions and indicating the set of heterogeneous tasks that
constitute the Simulation Workflow part (figure 1) [2].
Figure 1. Example of simple weather forecast workflow [2] that also includes a simulation workflow part.
Workflow description:
1. In steps 1-3: The weather is predicted for a particular geological area. Hence, the workflow is fed
with a model of the geophysical environment of ground, air and water for a requested area. Over a
specified period of time (e.g. 6 hours) several different variables are measured and observed. Ground
stations, ships, airplanes, weather balloons, satellites and buoys measure the air pressure, air/water
temperature, wind velocity, air humidity, vertical temperature profiles, cloud velocity, rain fall, and
more. This data needs to be collected from the different sources and stored for later access.
2. In steps 4-8 (simulations workflow): The collected data is analyzed and transformed into a common
format (e.g. Fahrenheit to Celsius scale). The normalized values are used to create the current state
of the atmosphere. Then, a numerical weather forecast is made based on mathematical-physical
models (e.g. GFS - Global Forecast System, UKMO - United Kingdom MOdel, GME - global model
of Deutscher Wetterdienst). The environmental area needs to be discretized beforehand using grid
cells. The physical parameters measured in Step 2 are exposed in 3D space as timely function. This
leads to a system of partial differential equations reflecting the physical relations that is solved
numerically. The results of the numerical models are complemented with a statistical interpretation
(e.g. with MOS - Model-Output-Statistics). That means the forecast result of the numerical models is
compared to statistical weather data. Known forecast failures are corrected. The numerical post-
processing is done with DMO (Direct Model Output): the numerical results are interpolated for
specific geological locations. Additionally, a statistical post-processing step removes failures of
measuring devices (e.g. using KALMAN filters).
3. In steps 9-10: The statistical interpretation and the numerical results are then observed and
interpreted by meteorologists based on their subjective experiences. Finally, the weather forecast is
visualized and presented to interested people.
A second more complex example illustrating the human and simulation workflows optimization in aerospace
SIMULATION
WORKFLOW
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
manufacturing industry is given next [3]. The manufacturer of an existing commercial passenger aircraft receives
feedback from customers requesting “a more comfortable cabin environment for the passengers”. The customer
request is then translated to the correlated specific alterations and test areas. Tasks and roles are defined, then
passed on to workflows optimization. The end result of optimal task/roles selection/configuration/schedule is
then executed. The general request is translated into a multi-disciplinary optimization problem that involves
alterations and tests in design areas such as:
• The passenger cabin environmental configuration: a CFD model (in-house)
• The passenger cabin structural configuration: an FEM model (in-house)
• The Environmental Control System (ECS): an electrical/thermal model (ext. partner A)
• The power plant (engines) configuration: an FEM model (external partner B)
• The human response evaluation: an ANN model (external partner C)
The necessary optimization loop involving the above models is described below and is shown in figure 2.
Figure 2. Example optimization loop involving heterogeneous simulation tasks in its workflow [4].
Workflow description:
The perceived comfort is evaluated by a Human Response Model (HRM), provided as a web-service by
external partner C. The HRM requires inputs of temperature, air flow, humidity and pressure (ENV) from
the Cabin CFD model and noise and vibration (N&V) inputs from the Cabin FEM model. The ENV
results of the Cabin CFD model are calculated based on its own operational characteristics in combination
with those of the ECS electrical/thermal model from external partner A. The N&V results of the Cabin
FEM model are calculated based on its own operational characteristics in combination with those of the
power plant FEM model from external partner B. Different settings and configurations for each of the four
models (FEM, CFD, electric/thermal) are considered based on the optimization loop algorithm and the
available constraints, the results of which are then evaluated by the HRM, leading to selection of new
values for calculation and evaluation. The loop continues until the selected criteria have been met.
In preparation for the human workflow optimization, specific tasks are determined for each model and the
overall optimization process loop, available related resources are catalogued and available related key personnel
are listed. The required simulations to be run are considered for generic settings and are not alterable.
A first schedule is produced as shown in figure 3 – upper part (for simplicity of the example, human operators
and computer resources are omitted. Also, minutes are depicted as days in the Gantt chart). This first schedule
then undergoes simulation workflow optimization. The relationships, requirements, costs and constraints of the
required simulations and the overall optimization loop are examined in detail by accessing and assessing the
information available from the semantic annotations of each model. These results are fed back into the human
workflow optimization to adjust the first schedule to the new, application-specific circumstances produced by the
simulation workflow optimization. A new schedule is produced that is now application-specific regarding
simulations (shown in figure 3 – lower part).
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
Figure 3. The first schedule from human workflow optimization (upper - orange) compared to the final schedule
after optimizing the simulation workflow & re-evaluating the human workflow (lower - green).
The Simulation workflow optimisation is necessary, in cooperation with the Human workflow optimisation, to
optimise the individual simulation parameters and configurations for the specific application, based on the overall
simulation tasks and requirements. The end result is to enable the application-specific optimisation of the
simulation tasks in the overall schedule produced by the Human workflow optimisation. The full concept
involves the production of several solutions by the SWO tool for the administrator to select.
1.2 Simulation Workflow components
A workflow can usually be described using formal or informal flow diagramming techniques, showing
directed flows between processing steps. Components can only be plugged together if the output of one previous
(set of) component(s) is equal to the mandatory input requirements of the following component. Thus, the
essential description of a component actually comprises only in- and output that are described fully in terms of
data types and their meaning (semantics). The algorithms' or rules' description need only be included when there
are several alternative ways to transform one type of input into one type of output - possibly with different
accuracy, speed, etc. Especially when the components are non-local services that are invoked remotely via a
computer network, like Web services, additional descriptors like Quality of Service, availability, etc. have to be
considered, too.
Single processing steps or components of a workflow can basically be defined by three parameters:
1. input description: the information, material and energy required to complete the step,
2. transformation rules, algorithms, which may be carried out by associated human roles or machines,
or a combination,
3. output description: the information, material and energy produced by the step and provided as input
to downstream steps.
2 THE DISTRIBUTED PDP FRAMEWORK
In the manufacturing domain, the Product Development Process (PDP) contains several stages that can be
improved by optimization. Design optimization is a very important stage as it reduces remanufacturing costs and
subsequent delays. Sustainability is another optimization task that reduces the environmental impact to the
product. Product testing and verification procedures also require optimization techniques in order to achieve the
most efficient schedule of both simulated and physical tests required (figure 4a).
As also shown in the second example above, today’s Product Development Processes (PDPs) are becoming
more and more decentralized and distributed. The PDP optimization problems are also more complex, with
multiple & contradictory objectives and they require powerful and/or specially designed optimization tools [5].
As a result, the corresponding human and simulation workflows are also becoming more complex as well as,
distributed and heterogeneous.
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
Under that scope, the EC project “Integrated management of product heterogeneous data – iProd” [6] aims to
improve the efficiency and quality of the PDP. This improvement involves the development and application of
test planning and optimization methodologies, which are part of the iProd Reasoning Engine, their end result
being detailed optimal workflows for applications areas such as Aerospace, Automotive and Appliances.
In this work we focus on the attempt to improve and optimize a distributed PDP and especially on
implementing an optimization method, in a distributed and heterogeneous network of collaborating systems and
tools. The aim is to present a flexible web based tool [7] that will able to promote a simulation workflow
optimization method, make it available to a remote application or another service and thus support a wider
automated collaboration between heterogeneous design & simulation tools (figure 4b).
(a) (b)
Figure 4. a) The SWO tool in the PDP improvement framework, and, b) the SWO tool as a web service.
3 THE SIMULATION WORKFLOW OPTIMISATION (SWO) TOOL
3.1 SWO Main Algorithm
The Simulation Workflow Optimization (SWO) tool is based on heuristic optimization techniques (Genetic
Algorithms) and delivers an optimized workflow implementation of the initial plan or schedule. The Genetic
algorithm is a class of Evolutionary Algorithms that works on the principle of survival of the fittest via natural
selection [8]. GA optimizers have been found to be much better than local optimization methods at dealing with
solution spaces having discontinuities, constrained parameters, and large no. of dimensions with many potential
local maxima. Evolutionary genetic algorithm (GA) optimizers are particularly effective when the goal is to find
an approximate global maximum in a multi objective optimization problem.
The SWO Genetic Algorithm performs the following steps:
1. Generates an initial population.
2. Computes the fitness for each individual.
3. Selects the parent couples
4. Creates the kids from the parents.
5. Selects the final members of the next generation
6. Returns to step 2 until a satisfactory solution is obtained.
The GA optimizer can use various forms of selection, cross-over/ mutation (steps 3 to 5) to evolve the initial
population. The important parameters of a GA are the: Population Size, Number of Generations,
Crossover/Mutation types & rates and Selection procedures, where:
• Crossover, is an exchange of substrings denoting chromosomes, for an optimization problem,
• Mutation, is the modification of bit strings in a single individual, and
• Selection is the evaluation of the fitness criterion to choose which individuals from a population
PRODUCT
DEVELOP
MENT
PROCESS
(PDP)
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
will go on to reproduce.
As shown in figure 1, the GA cycle (steps 2 to 6) is repeated until a termination condition has been reached,
such as: a solution that meets the criteria, the maximum number of generations, the maximum time allowed, etc.
The member of the last generation with the highest score(s) is the best solution and may be accompanied by the
other top candidates to create a set of best solutions proposed by the algorithm.
Figure 5. The SWO Genetic Algorithm flowchart
3.2 SWO Role in the PDP Improvement
In order to invoke the SWO tool a number of simulation task, application domain or product specific data are
required. SWO is using this information to adjust its functionality to the specific case under consideration. This
guarantees an improved performance as the tool can be specialized per domain, product or application by using
adequately prepared Knowledge Bases (KB) and ontologies that could provide the required information.
The main flow of operations of SWO is as follows:
1. Select the industry domain & product
2. Populate a KB with all necessary data
3. Select the PDP under revision and set the user requirements
4. Run any existing Work Breakdown & Task Planning tools to prepare the initial Human Workflow (HWF)
5. Isolate the Simulation Workflow (SWF) of interest (from inside the HWF)
6. Call the Simulation Workflow Optimization (SWO) tool for each heterogeneous task in the SWF
6.1 Retrieve from KB all simulation tasks & subtasks details
6.2 Create a population of candidate task schedules and calculate fitness & scores
6.3 Run the GA for a new generation
6.4 Select the fittest and Repeat fro step 6.2 until converged to the best schedule
7. Select the final/best improvements proposed – if insignificant the optimization has finished (step 9).
8. Submit the revised/new task information to KB and re-run the whole process from step 4.
9. Continue with the execution of the optimized workflow.
All required parameters about the PDP under consideration are retrieved from the KB and they define in detail
each simulation task and its subtasks.
A sample of the Genetic Algorithm convergence after a 80 generations is shown in figure 6, and a snapshot of the
user interface that controls SWO is shown in figure 7.
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
Figure 6. GA convergence to a final selection of individuals
Figure 7. A snapshot of the Web GUI that invokes the SWO tool.
4 CONCLUSIONS
In this document, the concept and the algorithm of the Simulation Workflow Optimization tool developed in
the iProd project was presented. The Simulation workflow optimization is necessary, in cooperation with the
Human workflow optimization, to optimize the individual simulation parameters and configurations for the
specific application, based on the overall simulation tasks and requirements. The end result is to enable the
application-specific optimization of the simulation tasks in the overall schedule produced by the Human
workflow optimization. The tool was designed to function in a distributed and heterogeneous environment that
usually describes today’s production environments.
Tsahalis J., Moussas V.C., and Tsahalis H.-T.
5 ACKNOWLEDGMENTS
The work presented in this paper has been performed under the EU-funded R&D project IPROD with
contract number FP7 FP7-ICT-2009-5-257657.
REFERENCES
[1] Webster’s Online Dictionary, http://www.websters-online-dictionary.org/
[2] Simtech Cluster of Excellence, University of Stuttgart, http://www.iaas.uni-
stuttgart.de/forschung/projects/simtech/sim-workflows.php
[3] Tsahalis, J.1, Tsahalis, H.-T.1 Moussas, V.C. (2012) “Modeling The Comfort Of Aircraft Passengers As Part
Of The Passenger Cabin Environmental Control System (ECS)”, Proceedings of International Conference
from Scientific Computing to Computational Engineering, 5th IC-SCCE, Athens, 4-7 July, 2012.
[4] Tsahalis, J.1, Tsahalis, H.-T.1 Moussas, V.C. (2013) “Optimization of a Heterogeneous Simulations
Workflow”, Proceedings of the 5th IC-EpsMsO, Athens, 3-6 July, 2013.
[5] Lee H. and Kim S.-S. (2001), “Integration of Process Planning and Scheduling Using Simulation Based
Genetic Algorithms”. Int J Adv Manuf Technol 18:586–590, 2001.
[6] iProd project: Integrated management of product heterogeneous data, http://www.iprod-project.eu
[7] Lau, R. (2007). "Towards a web services and intelligent agents-based negotiation system for B2B
eCommerce." Electronic Commerce Research and Applications 6(3): 260-273.
[8] Whitley L. Darrell (Ed) (1993), “Foundations of Genetic Algorithms”, MorganKaufmann 1993.
