Remote VisualizationRemote VisualizationUsing Low-Level Event ServicesUsing Low-Level Event Services

and Cluster Computingand Cluster Computing ——OROR——

ECho, cluster computing, ECho, cluster computing, and the Access Gridand the Access Grid

M. D.Wolf,

K. Schwan, and G. Eisenhauer

Georgia Institute of Technology

Members: Karsten Schwan, CS

Kevin Brennan, ECE

Uzi Landman, PHYS

Sudharkar Yalamanchili, ECE

Ellis Johnson, ISYE

David McDowell, ME

Suresh Menon, AE

http://www.cc.gatech.edu/projects/ihpcl

+ Research Scientists and

Graduate Students

transport model

...

chemical model

VisAD 3dActive Interface

3d vis &steering

Active Filter

Spectral2Grid

Spectral2Grid

historical chemical model results

heterogeneous client interfaces

or portals

information transformation and

routing

Collaborative HPC Applications

IHPCL

Access Grid

Undamaged

Damaged

Undamaged

vs.

W / and L /

interatomic spacing

L

W

L

W

Research Application: Molecular Dynamics

• Mechanical Engineering

• Physics

• Chemistry

•Aerospace Engineering

Real-Time, “Photographic” Visualization

Taken from the same research on the cover of the August 18, 2000 edition of Science.

Raytracing(pvmpov)

MolecularDynamics

Disk (binary format)every timestep

MediaConversion

EventChannel

Current Implementation

Viewer

VRML generation

AG

Raytracing(pvmpov)

MolecularDynamics

Disk (binary format)every timestep

EventChannel

What We Would Like

User Portal

User Portal

User Portal

EventChannel

(3D)

So What’s This Portal?

DirectoryServices

DataVideo

Vid

eo

Aud

io

Dis

trib

uted

P

ower

Poi

nt

local storage

Data Channels

Video Video

High Bandwidth

Data

• labeling• synchronization• replay• user-defined transformation• high performance

• handle “large” data feeds

Data Channels have

Event Channel

EventSink

EventSink

EventSink

EventSource

EventSource

Network Delivery

ECho Infrastructure

• Event Services for data and control structures• Interoperability

– Automatic conversion of data types between platforms

– Bindings for C and Java; Fortran almost completed

• No centralization, channels are represented by distributed data structures

• Local event delivery by subroutine call or queuing• Remote delivery multiplexed over shared

communications links (TCP, UDP, Wireless, multicast)

Message size

Ban

dw

idth

(M

bp

s)ECho Performance

0

10

20

30

40

50

60

100K 10K 1K 100b

EChoCORBAXMLMPI

ECho latency

0

50

100

150

200

250

Latency for 100Kb message (millisec)

EChoCORBAXMLMPI

Event Channel

NewEvent

ChannelF( )

Source Sink

Derived Event Channels

Derived Channels extend the ECho Model

• F() is a user-specified function, with on-demand, inline binary code generation • F() can be dynamically linked into the source• Next: Third party evaluation of F()

Event Channel

Event Channel

F1()Event

ChannelF2()

Event Channel

F3()

Source Sink

Chaining Derivations

Chaining Derivations

Source Sink

F1()

F2()F3()

SinkSink

AG

Raytracing(pvmpov)

MolecularDynamics

The Next Step...

local

F1()

F2()

F3()

3D+AGPortal

AGPortal

Data w

/ster

eo

Dat

a fo

r 2D

clie

nts

Color-coded

Color

-cod

ed

ECho References

• The ECho Event Delivery System, Greg Eisenhauer

• A Middleware Toolkit for Client-Initiated Service Specialization, Greg Eisenhauer, Fabian Bustamente and Karsten Schwan, Proceedings of the PODC Middleware Symposium - July 18-20, 2000

• Event Services for High Performance Computing, Greg Eisenhauer, Fabian Bustamente and Karsten Schwan, Proceedings of High Performance Distributed Computing (HPDC-2000)

• JECho - Supporting Distributed High Performance Applications with Java Event Channels, Dong Zhou, Karsten Schwan, Greg Eisenhauer and Yuan Chen, To appear at Cluster2000

Downloadable versions available from http://www.cc.gatech.edu/systems/projects/ECho/