A Pure Java HPCC Grid Architecture

01.05.2014 19:30

JUST.GridIACFeb2004.pdf (3,7 MB)

 

 ©2004 University of Applied Sciences Wolfenbuettel

and CLE, Department of High Performance Computing
LOGISTIK ZENTRUM
CLE
JUSTGrid
A Pure Java HPCC Grid Architecture
for Multi-Physics Solvers
Using Complex Geometries.
Jochem Hauser*, Thorsten Ludewig*
Torsten Gollnick*, and Hans-Georg Paap†
*Dept. of High Performance Computing
Center of Logistics and Expert Systems (CLE) GmbH
Salzgitter, Germany
†HPC Consultant
Barbing, Germany
presented at
IAC Rome, Italy, 2 February 2004
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 2
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Overview
Why Java for HPCC?!
What is JUSTGrid? (Framework)
JUSTGrid Communication Overview
Object Oriented Programming (OOP)
Threads
Graphical Applications
Java Performance Results
Conclusions
Future Work
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 3
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Java as the Language for HPCC
platform independence
simple and straight forward parallelization
unique included network capabilities
JDBC (Java Database Connectivity)
RMI (Remote Method Invocation)
Secure Connections over the Inter- and Intranet
easy generation of object reflecting the engineering design
process
,,code reusability'' - simplifies code design
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 4
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Why we like to use Java for writing highquality
portable parallel programs?
pure object formulation (i.e. an object representation of a
wing, fuselage, engine etc. described by a set of classes
containing the data structures and methods for a specific
item)
strong typing
exception model
elegant threading
portability
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 5
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
What is JUSTGrid
This is an age of possibility, and IT is the driving force behind this
change that occurs on a global range.
High Performance Computing and Communications (HPCC) on a
global scale is the key of this new economy.
The need for accurate 3D simulation in numerous areas of
computationally intensive industrial applications, including the rapidly
evolving field of bioscience, requires the development of ever more
powerful HPCC resources for a computational Grid based on the
Internet.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 6
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
What is JUSTGrid
The Java language has the potential to bring about a revolution in
computer simulation. Using Java's unique features, a multidisciplinary
computational Grid, termed JUSTGrid, can be built
entirely in Java in a transparent, object-oriented approach.
JUSTGrid provides the numerical, geometric, parallel, and network
infrastructure for a wide range of applications in 3D computer
simulation thus substantially alleviating the complex task of software
engineering.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 7
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Scope of JUSTGrid
3D Complex Geometries
Parallelization
Dynamic Load Balancing
Internet
Collaborative
Engineering Outsourcing Interactive
Steering
System
Security
Solver Visualization
Navier Stokes
(fluid dynamics)
Maxwell
(electromagnetics)
Schrödinger
(quantum mechanics)
Surface
Conversion
Debugging
Results Session Tracking
JUSTGrid a framework for HPCC in engineering,
science, and life sciences
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 8
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Scope of JUSTGrid
A solver only needs to contain the physics and numerics
of the simulation task for a single block or a single
domain.
The solver does not need to know anything about the
geometry data or the parallelization.
It has a simple structure
The solver can be tested independently before its
integration
Replacing the default solver by one's own solver.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 9
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Communication Overview
RMI based
Network
Solver/Code Development
Visualization
Collaborative Engineering
Server
Session-ID
Session-ID
Session-ID
Session-ID
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 10
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Communication Overview
Client Server
Client
sends the data AND the java solver code to the
server and receives a unique session id to
identify the session on the server.
Server
provides a framework for multi physics solver
Database Server
interactive steering
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 11
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Communication Overview
Client Server
with the unique session id everyone can
connect to a specific session on the server
start / stop session
changing the conditions of the computation
visualization
debugging
collaborative work
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 12
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
OOP - Object Oriented
Programming
Abstract Data Types
The association between the declaration of a data type and
the declaration of the code that is intended to operate upon
variables of this type
Data hiding and encapsulation
Protecting the data of an object from improper modification
by forcing the user to access the data trough a method.
class name
+variable1: integer
+variable2: double
+method1
+method2
Encapsulation of
data in an Object
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 13
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
OOP Example
Programming in 'C'
Programming in 'Java'
struct my_date {
int day, month, year;
} date;
date.day = 32;
public class MyDate {
private int day, month, year;
public void setDay( int day ) {
... validation code ...
}
}
MyDate myDate = new MyDate();
myDate.setDay( 32 );
ERROR!
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 14
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Threads
an efficient way of parallelizing codes
What are Threads?
Multithreaded programs extends the idea of multitasking by
taking it one level further: individual programs (processes) will
appear to do multiple tasks at the same time.
Programs that can run more than one thread at once are said
to be multithreaded.
Each task is usually called thread which is the short form for
thread of control.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 15
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
What are threads?
Three Parts of a Thread
A virtual CPU
The code the CPU is executing
The data the code works on
CPU
Code Data
A thread or
execution context
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 16
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Why Threads are good for CFD
Threads as a general parallelization strategy for CFD
codes
Sophisticated dynamic load balancing algorithms on
shared-memory machines
Advanced numerical schemes in CFD, i.e. GMRES, do not
require the same computational work for each grid cell.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 17
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ClientGUI
Why we like to use simple graphical user interface for the
JUSTGrid?
Because for a non-Programmer it is to difficult to collect all
different parts needed for a JUSTGrid session into a Java
source text and compile it for himself.
It is easier to run a quick test case without falling into
common programming traps.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 18
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
JUSTGrid Simple Client GUI
 
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 20
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
GRX Tool
Online
Visualization
Interactive
steering
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 21
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
GRX Tool
Online video generation
Integrated video player
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 22
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
GRX 3D Tool
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 23
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Virtual Visualization Toolkit
(VVT / ShowMe3D)
The idea of ShowMe3D is to develop a light weight
application, which is easy to use, with a limited (but useful)
set of functionality.
visualization
Geometry (surface)
results of a computation
debugging
online client - server connection
e.g. boundary updates
surface converting
e.g. quads to triangles
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 24
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ShowMe3D: Motivation
This program is designed for all programmers of Solver
Objects.
Based on the online "view" into the Server it can be very
helpful for debugging.
it is NOT designed to provide complete post processing
like TecPlot or Ensight
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 25
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ShowMe3D (continued)
Today's implementation of ShowMe3D contains
Visualization
Geometry data
a tree view of the Java 3D scene graph
Surface conversion
for Alias Wave Front Objects only
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 26
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ShowMe3D: main
The main window contains all the GUI elements for the file
input / output and visualization options
load geometry
save geometry
shaded view
wire frame view
system properties
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 27
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ShowMe3D: file types
ShowMe3D can load 2D and 3D object data in different file
formats
Triangle
Plot3D
Plane 2D
Plane 3D
Volume 3D
Autodesk/AutoCAD DXF
AliasWavefront Objects
3D StudioMAX
LightWave
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 28
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
ShowMe3D: Application
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 29
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Java High Performance and
Communications Test Suite
In the following series of slides we will demonstrate Java's
amazing numerical performance gains obtained over the
last few years.
Java numerical performance now rivals or exceeds that of
C or C++ codes used in engineering.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 30
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Simple numeric Benchmark
Simple numeric benchmark on a Sun Microsystems
Enterprise 10000 with 64 UltraSPARC II CPUs
1 2 3 4 5 6 7 8 910
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36 37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
7000
7500
8000
8500
9000
0
8
16
24
32
40
48
56
64
72
Simple numeric without communication 10e11 iterations
Time in s
Speedup
Number of CPUs
Time in seconds
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 31
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Mandelbrot Benchmark
Mandelbrot Set dimension 7200 x 4800, max iterations
5000 running 400 threads on a Sun Microsystems
Enterprise 6000 with 28 processors.
This code tests the self-scheduling of threads.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
0,0
2,5
5,0
7,5
10,0
12,5
15,0
17,5
20,0
22,5
25,0
27,5
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
110%
speedup
efficiency
Number of CPUs
Speedup
DEMO
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 32
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Matrix Multiply
for comparison, we have a Java
and a C++-coded version of the
sequential block-matrix multiply
that does not use threads and
multithreaded Java and C++
version.
to compare floating point
performance for scientific
applications between C++ and
Java on the test machines
to measure parallel efficiency
of a multithreaded application
Exactly the following source was used
for both benchmarks. (C++ and Java)
// get start time here
for( n=0; n
{
for( i=0; i
{
for( j=0; j
{
for( k=0; k
{
c[i][j] += a[i][k]*b[k][j];
}
}
}
}
// get end time here
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 33
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Sequential Matrix Multiplication
Runtime (2GHz Pentium 4, 1GB Memory) 1 run 2 run 3 run 4 run 5 run 6 run 7 run 8 run
3,15 3,19 3,22 3,16 3,15 3,17 3,16 3,16
3,23 3,23 3,25 3,23 3,23 3,23 3,23 3,25
3,86 3,88 3,90 3,90 3,90 3,90 3,89 3,90
3,55 3,51 2,12 2,12 2,12 2,12 2,13 2,12
GNU g++ -O3 -mcpu=pentium4
-march=pentium4 -Wall (Version 3.3.1)
Intel icc -O3 -mcpu=pentium4 -march=pentium4
(Version 8.0)
Sun Java HotSpot Client VM
(Version 1.4.2_02-b03)
Sun Java HotSpot Server VM
(Version 1.4.2_02-b03)
A sequential (1 thread) matrix multiplication using a 30
times 30 matrix doing 10000 iterations on a single
processor Pentium 4 PC running Linux.
After the two warmup phases in the Sun Java HotSpot
Server VM. This runtime is about 1.5 times faster then the
compiled C++ binary.
Due to a Linker error we could not use the -fast option
with the Intel compiler.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 34
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Multithreaded Matrix Multiplication
Multithreaded matrix multiplication using a 100 times 100
matrix doing 10000 iterations with 400 threads on a 26
CPU Sun Microsystems Enterprise 6000.
Runtime time in s
1.1.8_14 516,94
1.2.2_08 38,97
1.3.0_03 Server 37,47
1.3.1_02 Server 21,69
1.4.0_01 Server 19,51
1.4.1_02 Server 17,31
C++ - GCC 26,65
C++ - Forte 6u1 17,26
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 35
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Multi-threaded Matrix Multiplication
Results a 100 x 100 matrix doing 10,000 iterations with 400
threads on the E6000 (26 CPUs)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
0,0
2,5
5,0
7,5
10,0
12,5
15,0
17,5
20,0
22,5
25,0
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
110%
speedup
efficiency
Number of CPUs
Speedup
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 36
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Euler 3D Comparison
As a reference sample to check the correct
communication (boundary update) of the JUSTGrid we
computed a 3D cone with the JUST Euler 3D solver and
CFD++
JUST Euler 3D
CFD++
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 37
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Conclusions
With JUSTGrid a modern, well structured, easy to use and extensible
framework for HPCC is provided.
The code developer is freed from dealing with complex geometries, dynamic
load balancing and inter block or domain communication.
A numerical framework for a system of hyperbolic conservation laws is
installed, based on the integral form of the conservation equations
The parallel efficiency is obtained if a sufficient number of threads and
sufficient computational work within a thread can be provided.
The execution speed of Java code has increased substantially over the last few
years and now rivals the speed of C and C++ codes. More is too be expected.
Further work will be needed, but we following Kernighan's rules Make it right
before you make it faster as well as Don't patch bad code, rewrite it, the latter
rule being the reason for a pure Java flow solver code.
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 38
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Future Work
Extending the JUSTGrid parallel layer to work with
distributed memory machines (Beowulf cluster). (e.g.,
JavaSpaces, JINI, Sockets)
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 39
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
Acknowledgments
This work is partly funded by the ministry of Sciences and
Culture of the State of Lower Saxony, Germany under
AGIP 1999.365 EXTV program.
We are particularly grateful to Sun Microsystems,
Benchmark Center, Germany for providing exclusive
access to a 28 CPU Sun Enterprise 6000 server.
We are grateful to Mr. Jean Muylaert for providing
information about the European eXperimental Test
Vehicle.
This work is part of the Ph.D. work of Thorsten Ludewig
©2004 University of Applied Sciences Wolfenbuettel and CLE, Department of High Performance Computing page 40
Jochem Hauser et al. IAC Rome, Italy, 2 February 2004
References
Ginsberg,M., Häuser,J.,Moreira, J.E.,Morgan R., Parsons, J.C., Wielenga, T.J.
Panel Session: future directions and challenges for Java implementations of numeric-intensive industrial
applications published in: Advances in Engineering Software, Elsevier, 31,2000, p.743-751
Häuser,J., Ludewig, T., Williams, Roy D., Winkelmann, R., Gollnick, T., Brunett, S., Muylaert, J.
A Test Suite for High Performance Parallel Java published in: Advances in Engineering Software, Elsevier,31,
2000, p.687-696
Häuser,J., Ludewig, T., Williams, Roy D., Winkelmann, R., Gollnick, T., Brunett, S., Muylaert, J.
A Test Suite for High Performance Parallel Java paper presented at 5th National Symposium on Large-Scale
Analysis, Design and Intelligent Synthesis Environments, Williamsburg, VA, October 12th to 15th, 1999
Häuser,J., Ludewig, T., Williams, Roy D., Winkelmann, R., Gollnick, T., Brunett, S., Muylaert, J.
NASA Panel Java Soundbytes paper presented at 5th National Symposium on Large-Scale Analysis, Design and
Intelligent Synthesis Environments, Williamsburg, VA, October 12th to 15th, 1999
Häuser, J., Ludewig, Th., Gollnick, T., Winkelmann, R., Williams, R., D., Muylaert, J., Spel, M.,
A Pure Java Parallel Flow Solver, published in: Proceedings of the 37th AIAA Aerospace Sciences Meeting and
Exhibit, AIAA 99-0549 Reno, NV, USA, January 11.-14., 1999.
Häuser J., Williams R.D, Spel M., Muylaert J., ParNSS: An Efficient Parallel Navier-Stokes Solver for Complex
Geometries, AIAA 94-2263, AIAA 25th Fluid Dynamics Conference, Colorado Springs, June 1994.
 

 

Taggar:

| | |

Tillbaka