My Ph.D. Research

Major Research

Computer Sciences

Mechanical EngineeringMechanical Engineering

Mechanical Analysis

Computer Sciences

Design Optimization

Graphics User Interface

Parallel Computation

PPrroobblleemm SSoollvviinngg EEnnvviirroonnmmeenntt

CCoommppuutteerr AApppplliiccaattiioonn

PPaarraalllleell // DDiissttrriibbuutteedd GGrraapphhiiccss UUsseerr IInntteerrffaaccee

NNeettwwoorrkk CCoommppuuttiinngg

SSyysstteemm--LLeevveell

OObbjjeecctt--OOrriieenntteedd ((OOOOMMPPII))

PPoorrttaabbllee PPrroottooccooll ((MMPPII))

IIPPCC TTCCPP // IIPP

EEnnggiinneeeerriinngg SSiimmuullaattiioonn

SShhaappee OOppttiimmiizzaattiioonn

AAuuttoommaattiicc DDiiffffeerreennttiiaattiioonn

CCoommppiilleerr TTeecchhnnoollooggyy

FFiinniittee EElleemmeenntt AAnnaallyyssiiss

lel Programming

Portable

MPI Library

un, SGI, PC Linux, DEC, IBM RS-000

Workstation Cluster

messages passingsockets, telnet, rlogin

UNIX

CPU

MemCPU

Mem

CPU

Mem

CPU

Mem

UNIX UNIX

UNIX

ortable

MPI (Message - Passing Interface) -- Portable Paral

MPI Library

Sun Solaris multi-processors, SGI Challenge, Windows NT, OS/2 SMP

Parallel Extension

UNIX, WindowsNT, OS/2

CPU CPU CPU CPU

Memory

Shared Memory Machines

shared access to main memoryread / write

Parallel Extensions NOT Compatible -- Software NOT

MPI Library

nCUBE, Intel IPSC, Intel Paragon,CM-5

Parallel Extension

UNIX

CPU

Mem

Distributed Memory Machines

messages passingsend / recv

S6

CPU

Mem

CPU

Mem

CPU

Mem

MPI Applications -- Parallel / Distributed Software P

Multithreaded Programming Graphics User Interface: X-Window Event + Communication Message (IPC, TCP/IP) Client / Server for Parallel Computation:

1. Client: Computing + Receiving Results

2. Server: Sending Results + Computing Job Scheduling – Task Management: Synchronization + Load Balancing

Parallel Computing without Task Manager

Parallel Computing with Task Manager JAVA Threads:

1. Synchronization (Image Loading)

2. Concurrency (Slide Show, Clock)

Shallow Water Equations Model (SWE)

Performance of Parallel Computation - ParaGraph

Performance of Parallel Computation - Upshot

n of Physical Parts

odeling

nternal iteration

Convergencestimate

alysis &ptimization

ral Analysis

ve Approachmesh density

no

yes

Postprocessing

Parallel ShapeOptimization

Design OptimumEstimate

yes

no

vity Analysis

An Abstract View of the EPPOD System for the Desig

Prototype

InitialDesign

Model Description

PATRAN CommandLanguage

ModelSelection

ImprovedDesign

Interactive Input

M

i

Error E

AnLocal O

Structu

Adapti- new

Global Optimization

yes

no

Parallel DesignSensitivity Analysis

Parallel ShapeOptimization

Design OptimumEstimate

Sensiti

external iteration

XXoX CSGLanguage

Preprocessing

ParallelOptimal Domain Decomposition

Parallel MeshGeneration

Mesh Refiner

ment

mesh & splitting

simulation

The Tools of a Problem Solving Environ

Geometry Specification

geometry specificationMesh Generation

Mesh Splitting

Structural Analysis

Shape Optimization

optimization

The Flow Diagram of the Implemented EPPOD System

FEA result

FEA result

cancel

FEA start

boundarymesh

ADS Computation

If Restart

Yes

No

No

No

No

No

Yes

Yes Yes

Pause

YesStop

FEM Pre-Processor

Mesh Generation

Domain Decomposition

(X-Windows Server)

FEM Processor

Displacement Analysis

Stress Analysis

Object Analysis

(X-Windows Server)

FEM Post-Processor

FEM mesh Display

FEM contour Display

Deformation Display

(X-Windows Server)

ADS Initialization

Stop

Stop

Start

If Interrupt

If Converge

If Error

If Pause

EPPOD - Electronic Prototyping for Physical Object Design Interactive System

FEM Pre-Processor Server

FEM Post-Processor Server

FEM Processor Server

UNION

ION ROTATE

Box - 1

er - 2

TRANSLATE

Cylinder - 3

Geometry Definition

XXoX: CSG (Constructive Solid Geometry) Language

cylinder1 = cylinder(0,0,-0.52,1.56,1.04)cylinder2 = cylinder(0,0,-0.52,1.04,1.04)cylinder3 = cylinder(0,0,0,0.5,5.4)box1 = box(-0.26,-0.52,1.04,0.52,1.04,6.6)rod = rotate(cylinder1-cylinder2,0,0,0,0,1,0,-90)rod = rod | box1 | translate(scale(rod,1,0.5,0.5),0,0,8.16)rod = rod | translate(rotate(cylinder3,0,0,0,0,1,0,-90),2.6,0,8.16)

CSG Tree:

DIFFER

ROTATE

TRANSLATEUN

UNION

Cylind

SCALE

Cylinder - 1

deling System

Viewing

Information

XXoX - An Interactive X-Window based Solid Mo

Menu

Operation

Message

Command

Drawing

ftwares

VRML

Rendering #3

Interfaces between XXoX and Foreign So

PATRAN

Rendering #1

Rendering #2

ftwares

XPoly

Rendering #3

Interfaces between XXoX and Foreign So

PATRAN

Rendering #1

Rendering #2

Virtual Reality Modeling Language (VRML) Interface

ethodology

esh of step ed to a finer arallel.

5. An optimal mesh splitting scheme to mini-mize the bisection width is applied.

Parallel Mesh Generation and Decomposition M

1. An adaptive mesh algorithm is invoked to gener-ate an initial “coarse” mesh.

2. A scheme to split the initial mesh into equal-sized subdomains is applied.

3. A linking rou-tine to form the new subdomain boundaries is called.

4. The malgorithm1 is appligenerate mesh in p

Windows of the GUI for Mesh Generation and Splitting

Triangular element mesh generation & element-wise domain decomposition

Triangular element mesh generation & node-wise domain decomposition

Quadrilateral element mesh generation & element-wise domain decomposition

Adaptive Ap

Deformed shape

Stress: <-7789.59, 6769.68>

Error: 0.164/2.10E-5

79 nodes, 117 elements

15 refine pointsin

one adaptive step

Deformed shape

Stress: <-16446.6, 8286.24>

Error: 1.427/8.68E-5


proach

Performance of Parallel Mesh Generation - Engine Rod

Utilization Count: states of idle, overhead, and busy as function of time.

Utilization Summary: overall cumulative percent-age of time in idle, overhead, and busy states.

Speedup - sequential : T1s / Tp.Speedup - parallel : T1 / Tp.

Engine Rod

Performance of Parallel Mesh Generation - Torque Arm




Torque Arm

Performance of Parallel Mesh Generation - Engine Axis




Engine Axis

Performance of Parallel Mesh Generation (Paragon) - Ex1d

SpeedUp : T1 / Tp & T1s / Tp

Performance of Parallel Mesh Generation (Paragon) - Ex1d


Performance of Parallel Mesh Generation (nCUBE) - Ex3a


Performance of Parallel Mesh Generation (nCUBE) - Ex3a


Mesh Decomposition and Sparse Matrix - Engine Rod

DFS - basicCommunication: 17 / 10 Bandwidth: 4 / 24Connectivity: 51 IBV: 28

BFS - domain-wiseCommunication: 8 / 4 Bandwidth: 7 / 15Connectivity: 16 IBV: 15

BFS - strip-wiseCommunication: 15 / 8 Bandwidth: 9 / 16Connectivity: 30 IBV: 22

Eigenvector SpectralCommunication: 7 / 5 Bandwidth: 7 / 22Connectivity: 14 IBV: 12

Polar - recursive bisectionCommunication: 16 / 11 Bandwidth: 2 / 6Connectivity: 48 IBV: 29

Inertia - first eigenvectorCommunication: 12 / 6 Bandwidth: 7 / 17Connectivity: 36 IBV: 19

Polar - local optimumCommunication: 8 / 5 Bandwidth: 5 / 11Connectivity: 16 IBV: 15

Cartesian - local optimumCommunication: 7 / 4 Bandwidth: 6 / 15Connectivity: 14 IBV: 14

Mesh Decomposition and Sparse Matrix - Engine Cap









Mesh Decomposition and Sparse Matrix - Engine Axis









Performance of Mesh Decomposition

Maximum Interface Length

BandwidthInterpartitioning Boundary Vertices

Subdomain Connectivity

Computation Time

Methods

CLO Cartesian Local OptimumPLO Polar Local OptimumCLE Cartesian Longest ExpansionMRSB Multilevel Recursive Spectral BisectionGreedy Domain-Wise BFSRCM Strip-Wise BFSInert Inertia - First Eigenvector

e Optimization

Iter# 5,6 (eval# 81,97)

Iter# 7-10 (eval# 100-122)

Iter# 4 (eval# 65)

Numerical Performance for the Traditional Shap

Iter# 1 (eval# 17)

Iter# 0 (Initial Design)

Iter# 3 (eval# 49)Iter# 2 (eval# 33)

Traditional Shape Optimization

tion Method

Local Iter# 3 (eval# 15/51)

Global Iter# 3

Global Iter# 2

Numerical Performance for the Model Coordina

Local Iter# 1 (eval = 5/17)

Iter# 0 (Initial Design)

Local Iter# 2 (eval# 10/34)Global Iter# 1

Model Coordination Shape Optimization

Parallel Electronic Prototyping - Engine Rod


Shape OptimizationZ = Area min.dispalcement 0.00050.1 Xi-j 1.0 i = 1, 2

j = 1, 20.5 X3-j 1.0 j = 1, 20.5 X4-j 1.0 j = 1, , 7

Deformed shape

X1-1

X4-2

X1-2

X2-2

X3-2

X4-6

X4-5X4-3

X3-1

X2-1

X4-4

X4-1 X4-7


Displacement

Stress

Deformed shape

Displacement

Stress

Parallel Electronic Prototyping - Torque Arm


Shape OptimizationZ = Area min.dispalcement 0.0015stress 100000.1 Xi-j 1.0 i = 1, 2

j = 1, 20.5 Xi-j 1.0 i = 3, 4

j = 1, 20.5 X5-j 1.0 j = 1, , 50.2 Xi-j 1.2 i = 6, 7

j = 1, 20.35 X8-j 0.6 j = 1, 20.35 X9-j 1.0 j = 1, , 50.2 R1 = X10 0.450.15 R2 = X11 0.275

Deformed shape

X1-1


Displacement

Stress

Deformed shape

Displacement

Stress

X2-1X3-1

X4-1

X5-3X5-2X5-1

X1-1

X5-4

X5-5

X4-2X3-2X2-2X1-2

X6-2

X7-1

X6-1

X7-2

X8-2X9-1 X9-5

X9-3X9-4X9-2

X8-1

R1=X10

R2=X11

Performance of Parallel Shape Optimization - Engine Rod

Speedup - parallel : T1 / Tp.

Engine Rod

Utilization Countstates of idle, overhead, and busy as function of time.

Utilization Summaryoverall cumulative percentage of time in idle, overhead, and busy states.

proc#=4

proc#=16

Performance of Parallel Shape Optimization - Torque Arm

Speedup - parallel : T1 / Tp.

Torque Arm

Utilization Countstates of idle, overhead, and busy as function of time.

Utilization Summaryoverall cumulative percentage of time in idle, overhead, and busy states.

proc#=4

proc#=16

Finite Element Analysis Post-Processor - Xcontour

Result of Structural Stress

Contour Lines

Numbering of FEM Mesh

Deformed Shape

My Ph.D. Research

Technology

parallel extensions

task managerparallel

mpi message

tcpip client server

windows nt

portable shared access

eppod electronic prototyping

task manager java threads