This document discusses using the Service-ORiented Computing EnviRonment (SORCER) platform for deterministic and stochastic optimization problems. It describes how global optimization algorithms VTDIRECT95 and QNSTOP were modified to run as optimization services on SORCER. As a case study, the EBF3PanelOpt framework for structural optimization of curvilinear aircraft panels was integrated with SORCER. This allowed the optimization problem to be solved in a distributed manner across computational resources for significantly faster objective function evaluations.
5. Optimization Algorithms
Optimization algorithm – a procedure which is executed
iteratively by comparing various solutions till an
optimum or a satisfactory solution is found.
VTDIRECT95 and QNSTOP:
Global optimization algorithms well suited for
multidisciplinary design optimization (MDO).
Exhibit parallelism.
Able to accommodate problems of higher dimensions.
6. Objective Function
An equation to be optimized given certain constraints
and with variables that need to be minimized of
maximized using nonlinear programming techniques.
Point co-
ordinates (X)
Objective
Function
F(X)
8. Motivation
Requirements of conceptual design of complex
systems – extensive exploration of design space and
analyses of a large number of potential design
configurations.
Traditional conceptual design – low fidelity models,
poor accuracy.
Physics based modeling – better accuracy,
computationally intensive.
Need for a platform to cope with computational
complexity, high design time and high cost of
production.
9. SOC and SORCER
Addresses the challenges faced by HPC in terms of
scalability, availability, reliability, and flexibility.
Service-oriented computing (SOC) – utilizes platform-
agnostic services that effectively communicate with one
another to perform user-requested tasks in a distributed
computing environment.
Service
Registry
Find
Service
Requestor
Publish
Bind
Service
Requestor
Service
Provider
10. SOC and SORCER (cont.)
SORCER – derived from the SOA (service-oriented architecture)
model; Java-based, network-centric computing platform that
enables execution of service-oriented programs.
Advantages:
Large scale, distributed, decentralized
Leveraging the power of HPC
Reusability
Cost effective
Better utilization of computational resources
Load balancing across computational resources
11. Modifying VTDIRECT95 and
QNSTOP for SORCER
Use of JNI (Java Native Interface) to allow Java
applications to invoke native code and vice versa.
Use of JNI’s invocation interface – allows a regular
non-Java program running on the native operating
system to invoke a JVM to gain access to Java classes
and features.
JNI wrapper – layer of abstraction between the
optimization algorithm and the Java block that
evaluates the objective for a given design point.
13. VTdirect and QNSTOPS as
SORCER services
Leveraging the power of Exertion-oriented Programming
(EOP) to make a number of services available to users in
a distributed computing environment.
SORCER terminology:
Service provider: A remote object accepting exertions
from service requestors and performs calculations.
Service requestor: An object that creates exertions and
submits them to the grid.
Exertion: Defines collaboration – service-oriented
programs.
14. VTdirect and QNSTOPS as
SORCER services (cont.)
Steps involved in creating a provider that allows use of
VTdirect as a service:
Wrap the Fortran 95 subroutine with JNI.
Set up infrastructure for a SORCER provider.
//Code to execute the corresponding executable
./vtdirect
Set up infrastructure for a SORCER requestor.
//Exert collaboration
Exertion result = vtdirectTask.exert();
15. Engineering Application: Optimization of
Curvilinear Blade-stiffened panels
EBF3PanelOpt Framework – structural optimization of
curvilinearly stiffened panels to facilitate building of
optimal light-weight structures.
Deployed as a provider within SORCER.
24. Conclusion
The algorithms packages VTDIRECT95 and QNSTOP were
successfully implemented as SORCER services.
The EBF3PanelOpt framework was successfully integrated with
SORCER, facilitating the optimization of curvilinearly stiffened
panel in a truly distributed manner.
With the use of SORCER’s JavaSpaces technology, computers
could be added on the fly, hence providing flexibility.
Distributed parallelism and load balancing across
computational resources with SORCER helped in significantly
speeding up objective function evaluations in a dynamically
scalable environment.
26. References
[1]. Deshpande S., Watson L. T., Love N. J., Canfield R. A.,
and Kolonay R. M., “Aircraft Design Markup Language for
Multidisciplinary Aircraft Design and Analysis”, AIAA –
Journal of Information Sciences, Vol. 12, No. 2, Feb. 2015.
Editor's Notes
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
An effective global method that avoids being trapped at local optima and intelligently explores potentially optimal regions to converge globally for Lipschitz continuous optimization problem.
Advanced features – derived data types, pointers, dynamic memory allocation – design dynamic data types that flexibly organize data on a single machine, effectively reduce local storage and share data among multiple processors
Balancing global and local search --- do a little bit of both in every iteration.
Choosing rectangles that have the lowest lower bound for some rate of change constant.
Small constants select rectangles good for local search; large constants select rectangles good for global search.
Higher dimension problems
No need of a Lipschitz constants
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
Aircraft design – multidisciplinary process – involves several disciplines --- contribute to achieving an optimal design; satisfying all requirements.
Conceptual design – first – study of a large number of design configurations and what impact these designs have on the performance
SOC – HPC + orchestered services => agile applications that scale across platforms and organizations
Great flexibility – loose coupling of components
Service providers accept remote messages from requestors to execute a collaboration. These messages – exertion (specify relationship between services and how information is passed between them)
Exertion – service data, operations and control strategy
Collaboration – process; the specification of collaboration – exertion; dynamic federation of peers – implementation of collaboration.
Adv – network-centric messaging, fault tolerance
Capability to optimize flat/curved multi-sided panels with straight/curved edges having curvilinear blade type stiffeners under multiple loading conditions.
Capability to optimize flat/curved multi-sided panels with straight/curved edges having curvilinear blade type stiffeners under multiple loading conditions.
Capability to optimize flat/curved multi-sided panels with straight/curved edges having curvilinear blade type stiffeners under multiple loading conditions.
A rectangular panel of size 0.4064 m * 0.5080 m representing a large wing engine pylon rib.
Subjected to combined compression and shear in-plane loads.
Buckling constraint – close to 1. corresponding masses are optimal.
KSC criteria – panel is getting more stressed for better optimal designs
Crippling criteria – stiffener is more prone to crippling or local failure
(one or more flanges buckle in a local buckling mode with wavelength unrelated to the length of the beam)
Taskers – accept exertion tasks
Jobbers – manage service collaboration for PUSH
Spacers – manage service collaboration for PULL
Contexters – provide data contexts
Catalogers – QoS
Exert Monitors – monitor execution of running exertions
Persisters – persist data context, tasks and jobs to be reused for EO programming