grid everywhere - jpgrid.org · maestro: nasa jet propulsion laboratory first steps towards the...

Adrian Cockcroft

Chief Architect

High Performance & Technical Computing

Grid EverywhereAccelerating Innovation: Sun's HPTC/Grid Strategy

Grid Everywhere

www.sun.com/solutions/hpc/

Maestro: NASA Jet Propulsion LaboratoryFirst Steps Towards the Inter Planetary Grid� Problem

Simulate remote control and data collection

Global web access to image data

Multi-platform visualization

� Solution Multi-site, multi-organizational collaboration

Open standard imaging technology

JAVA Advanced Imaging and JAVA 3D

� Result Universal access to scientific data

Real time public involvement in science and exploration

NASA recorded 109 million hits in the first 24-hours

250,000 Maestro software downloads in 3 weeks

"I love this program. I have never felt so close to space exploration as I do when I'm poking around it. It is an awe inspiring mission and this software practically lets you touch it."

E-mail to JPL commenting on Java 3D& JAI based Maestro software

AgendaAgenda

HPTC @ SunHPTC @ SunEarly AdoptersEarly Adopters

Grid TechnologiesGrid TechnologiesGrid Alliances and StandardsGrid Alliances and Standards

Grid SolutionsGrid SolutionsVisual GridVisual Grid

� Perform mathematical modeling of natural phenomena (science and engineering) or business processes (business optimization)

Crash Simulation, Risk Management, Electronic Design, Gene Modeling, Weather forecasting, etc.

� Many possible types of work loads

Transactions, Queries, Numerical, Visual/Multi-media, Networking

� Represent EARLY ADOPTERS, often a leading indicator of future IT requirements. (64bit, SMP, Web, Grid, graphics,....)

What does HPTC mean @ Sun?

8000 Grids using Sun technology350,000 CPUs served

Sun: HPTC Volume Leader

� Sun is the volume leader based on IDC Data for 2003

� CY2003: Technical server shipments 89,907 totalSun #1 with 40.2% market share (36,156 units)

� CY2003: Technical server revenue $5.4B totalSun #3 with 18.3% market share ($981M)

IDC Technical Server Tracker February 23rd 2004

Q402 Q103 Q203 Q303 Q403

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

11000

12000

13000

14000

Technical Server Unit Shipments

Q302 Q402 Q103 Q203 Q303 Q403

0.0

25.0

50.0

75.0

100.0

125.0

150.0

175.0

200.0

225.0

250.0

275.0

300.0

325.0

Technical Server Revenue $M

Early AdoptersEarly Adopters

Centerof Excellence

Proof ofConcept

Reference Architectures and Solutions

CustomerReady

Systems

AdvancedSystems &Services

Solutions DevelopmentMethodology

Productsand Labs

Technology

Sun HPTC Transforms Innovation into Customer Value

Early Adopter Example: Interval MathCoordinate research, partner and take it to market

Interval ComponentsInterval is a span of all numbers between two values [x, y]Interval math manages numerical accuracy and stability issuesSun Fortran has Interval Datatype and operationsSolver libraries are being researchedInterval algorithms existMore scalable, and more efficient (fewer GF to solve)

The Interval ProblemMost textbooks say you can't solve nonlinear optimization

problems true for point solutions only!!

Key Book: Global Optimization Using Interval Analysisby Eldon Hansen, G. William Walster ISBN: 0824740599

Hill Climb Optimization ExampleTry hill climb algorithm on a difficult dataset, with very narrowand specific optimal solutions it's not a robust solution....

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Telegraph poles on the prarie

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Sample Prarie

Interval Optimization ExampleInterval solver tells you the range of the possible solution overan interval, here we split the data into four intervals and evaluate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Telegraph poles on the prarie

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Prarie Interval ranges

Interval Optimization ExampleKeep zooming in until the result is a small enough interval, and youend up with a deterministic and correctly bounded solution

00.05

0.10.15

0.20.25

0.30.35

0.40.45

0.50.55

0.60.65

0.70.75

0.80.85

0.9

Zoom in on Maximum

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Interval Driven True Solution

The Answer!

Grid TechnologiesGrid Technologies

The Strategic Need for Shared Resources

� Life Sciences

� Electronic Design

� Financial Services

� Automotive Manufacturing

� Scientific Research

� Oil and Gas Exploration

� Telecommunications

� Business Computing

Discovery Informatics, database queries

Grid enabled enterprise applications, database and transaction processing

Simulations, verifications, regression testing

Enhanced delivery of network services

Simulations, seismic analysis, visualization

Large computational problems, collaboration

Crash testing simulations, stress testing, aerodynamics modeling

Risk and portfolio analysis, simulations

Industries Computing Tasks

Types of Grid

SensorGrid

DesktopGrid

ClusterGrid

EnterpriseGrid

GlobalGrid

DataCapture

Computation andVisualization

DataStorage

Trade Exchange

VisualGrid

The Value of Grid

The business benefits of Grid are those of a network-wide consolidation (across line of business boundaries and geographies)

Enable a service-oriented, automated delivery of IT

Improved utilization

� Enables new capabilities and services� Lower operational costs of running IT

Maximize productivity through optimal use of

horizontally or vertically scaled systems

Sun's Evolutionary Compute Grid Strategy:Expand existing capabilities and integrate new ones

Cluster GridResource Sharing

Simplest Grid deployment Optimal utilization of resources Enables new capabilities and services Lower operational costs for IT Maximize productivity

Enterprise GridJava Web Services, Data Grid

Resources shared within the enterprise Policies ensure computing on demand Gives multiple groups seamless access to

enterprise resources Meets commercial enterprise standards

Global GridTrade Exchange

Resources shared over the Internet Global view of distributed resources and data Trade Exchange for compute capacity and

data access Foundation for Utility Computing

Grid: Where is your reality today?What is the effect upon your business?

No grid - Local grid Enterprise grid Global grid

Failing Surviving Profiting Investing

Old Reality Current Reality Future Reality

Lagging - Conforming Exploiting Leading Pioneering

Ch

asm

Grid Crossing the Chasm

Early adopters

HPTC has led in adoption of Grid technologies

Technical Computing

Clusters

GlobalGrid

EnterpriseGrid

Ch

asm

Grid Crossing the Chasm

Early adopters

HPTC methodologies and expertise are key for Grid to cross the business computing chasm

Technical Computing

Clusters

GlobalGrid

EnterpriseGrid

Cha

sm

Business Computing

Grid Datacenter

Enterprise Grid System Components

Virtualized Physical Resources

Java Web Services

Utility Computing

Bu

sin

ess

M

od

el

Tech

no

log

y Fo

un

dati

on

What it takes

� Thought Leadership Trade exchange model, Utility Computing, Reference Architectures,

Infrastructure Solutions, Grid Professional Services practice

DARPA HPCS High Productivity Computing System covers administration and developer issues, not just peta-scale performance

� Expertise HPTC experience to assemble solution recipes out of technologies

Professional services to architect, implement, manage

� Technology Java web services, infrastructure and tools (JES, Java Studio, OSS/J)

Automatic cluster configuration (Sun Control Station/SDSC ROCKS)

Optimal matching of workloads to resources (N1 Grid Engine)

Automated provisioning of resources, applications and services (N1 Grid )

Advanced platforms (SPARC, Opteron, Solaris, Linux, SAN/NAS, Visualization)

� Plus alliances and standards....

Alliances andAlliances and

StandardsStandards

Sun Grid Services Environment

GlobusToolkit

Grid Engine Portal

JAVAEnterprise

System

N1 Grid Engine

Web Service Enabled

Enterprise Grid Enabled C C++

FORTRANApplications

JAVA

Network Enabled

Global Grid Enabled

DRMAAStandardInterface

Alliances and Standards

� Partners Sun has wide ranging agnostic HW/SW/Services/Channel alliances

HPTC@Sun has signed up more than 25 interconnect and grid partners

� Global Grid Forum GGF - www.gridforum.org Sun led the creation of the GGF DRMAA standard

Sun holds an Area Director position within GGF

� Globus Alliance - www.globus.org Many Globus toolkit deployments are on Sun

Sun is working closely with Globus on future developments

� Enterprise Grid Alliance www.gridalliance.org Sun is a founding member of the EGA

The EGA is working to specify and standardize Enterprise Grid solutions

� DMTF and OASIS Sun has many representatives and is working to make sure that future Grid

standards are truly open and royalty-free

Enterprise Grid Alliancewww.gridalliance.org

� Focus Practical Solutions for Enterprise Grid Example: run General Ledger on a Grid (not scientific research grid!)

Short term database cluster grid and datacenter virtualization

Medium term combine to run technical applications in spare cycles

Long term global grid integration once standards have matured

� Board-level Members EMC, Fujitsu-Siemens, HP, Intel, NEC, Network Appliance, Oracle, Sun

� Initial Working Group Proposals Reference Model vocabulary, context, use cases

Component Provisioning standardization

Data Provisioning standardized high level functionality

Utility Accounting use cases, data models, mappings to standards

Grid Security mapping to existing security standards

� Relationships Many EGA member companies and people are members of other groups

Adopt mature standards from GGF, DMTF, OASIS, TMF, SNIA, ITIL...

New! April 2004

Grid SolutionsGrid Solutions

SDSC ROCKstar Supercomputer0 to #201 in under 2 hours!

SDSC ROCKstar: A 128 node/256 CPU x86 cluster, #201 in Top 500SDSC ROCKstar: A 128 node/256 CPU x86 cluster, #201 in Top 500

Built live on the SC2003 show floor in 100 minutes, adding a new Built live on the SC2003 show floor in 100 minutes, adding a new

server every 30 seconds. 200 jobs were submitted to the queue server every 30 seconds. 200 jobs were submitted to the queue

and the cluster began to execute 15 minutes into the build cycle.and the cluster began to execute 15 minutes into the build cycle.

Grid Computing ComponentsBoth Horizontal Scale-out and Vertical Scale-up

Sun Fire E25K

Sun Fire E20K

Sun Fire E6900

Sun Fire E4900

Sun Fire E2900

Sun Fire V880

Sun Fire V480

Sun Fire V440

Sun Fire Blades

Sun Fire V20z (Opteron) Sun Fire

TM

V210

(SPARC)Sun BladeTM

Workstation

UltraSPARC

IA32, AMD64

UltraSPARC

Capability Computing

Services

CapacityComputing

Services

Prebuilt Racks

The Grid Architecture Dilemma:Large memory SMP nodes or many thin nodes?

Workload scales vertically:� Parallel codes: OpenMP� Large Shared Memory� Top Performance� Higher acquisition cost� Lower development and

management complexity & cost

Workload scales horizontally:� Serial and parallel codes: MPI� Optimized Throughput� Lower acquisition cost� Higher development and

management complexity & cost

Acquisition Cost / CPU

The DecidingFactor

What do theworkloads

require?

Complexity Cost / CPU

CMT On-Chipxxx GB/s0.1 - 0.01 µs$xxx?

Interconnect ComponentsMapping Out Bandwidth and Latency

Ethernet0.1 GB/s100 - 10µs$xxx

Myri/IB/Q/FL

0.4 – 4.8 GB/s10 - 1 µs$x,xxx

Memory

9.6 - 57 GB/s1 - 0.1 µs$xx,xxx

Latency(invertedlog scale)

10ns

100ns

1us

10us

100us

Proximity

SystemCall

LibraryCall

Load/StoreInstruction

0.1 1 10 100 1000Gigabytes/sec Bandwidth (on logarithmic scale)

Interconnect ComponentsMapping Out Bandwidth and LatencyEstimated/Approximate performance numbers

0.1 1 10 100 1000Gigabytes/sec Bandwidth (on logarithmic scale)

Latency(invertedlog scale)

10ns

100ns

1us

10us

100usGBESolaris 9

GBELinux

10GBELinuxAMD

UsermodeGBELinux

QuadrixLinuxAMD

Myrinet IB x4LinuxAMD

IB x12LinuxAMD

SMPV20z

SMPE25K

SMPE6900

SMPV440

CMTFuture

PCI-X 133SpeedLimit

MPI calllatency Limit

10GBEPCI-E

PCI-Expressx8 SpeedLimit

Visual GridVisual Grid

High End Visual System EvolutionTrend: Disaggregation and Virtualization of Components

Starting point: Applications run on single system with rendering Starting point: Applications run on single system with rendering

distributed across multiple 3D Graphics Accelerators and hardwired displaydistributed across multiple 3D Graphics Accelerators and hardwired display

3D Graphics3D GraphicsAcceleratorAccelerator

3D Grpahics3D GrpahicsAcceleratorAcceleratorXlibXlib

X11 ServerX11 Server

IntegratedIntegratedVisualVisualSystemSystem

OpenGLOpenGLServerServer

X11X11ServerServer

Xinerama, MDU, CAVElib, etc.Xinerama, MDU, CAVElib, etc.

DDXDDX OpenGL DPOpenGL DP

ApplicationApplication

OpenGLOpenGL

StorageStorageSystemSystem

SAN orNetwork

High End Visual System EvolutionVirtualize and Distribute 3D Display Capability

Virtualize Display: Applications run on single system with rendering Virtualize Display: Applications run on single system with rendering

distributed across multiple 3D Graphics Accelerators and Virtualized Displaydistributed across multiple 3D Graphics Accelerators and Virtualized Display


3D Grpahics3D GrpahicsAcceleratorAcceleratorXlibXlib


IntegratedIntegratedVisualVisualSystemSystem


X11X11ServerServer




OpenGLOpenGL


SAN orNetwork

SwitchSwitchDisplayDisplay

viaviaVideoVideoSwitchSwitch

ororRedirectRedirectover IPover IP

NetworkNetwork

IP

Video

High End Visual Grid SystemVirtualize and Distribute 3D Accelerators

Applications run on High Capacity Enterprise Server with rendering Applications run on High Capacity Enterprise Server with rendering

distributed via Grid Engine across a visual grid of multiple 3D server nodesdistributed via Grid Engine across a visual grid of multiple 3D server nodes


DZGLDZGL

XlibXlib

3D server3D server

3D Grpahics3D GrpahicsAcceleratorAccelerator

DZGLDZGL

XlibXlib

3D server3D server


DZGLDZGL

XlibXlib

Application ServersApplication Servers


X11X11ServerServer


X11X11ServerServer




OpenGLOpenGL

High Bandwidth Grid Interconnect


SAN orNetwork

IP

Video




NetworkNetwork

IP




NetworkNetwork

IP

High End Visual Grid SystemVirtualize and Distribute Application and Visualization

Applications virtualized using Grid Engine on a cluster of any kind of servers withApplications virtualized using Grid Engine on a cluster of any kind of servers with

rendering distributed across multiple 3D server nodesrendering distributed across multiple 3D server nodes


DZGLDZGL

XlibXlib

3D server3D server

3D Grpahics3D GrpahicsAcceleratorAccelerator

DZGLDZGL

XlibXlib

3D server3D server


DZGLDZGL

XlibXlib

Application Server ClusterApplication Server Cluster


X11X11ServerServer


X11X11ServerServer




OpenGLOpenGL

High Bandwidth Grid Interconnect


SAN orNetwork

IP

Video




NetworkNetwork

IP




NetworkNetwork

IP

The University of Texas at AustinVisualization at the ACES Visualization Lab

"The speed of graphics and superior image quality of the Sun Fire V880z visualization system will play a significant role in these key areas of technology and research."

Dr. Kelly Gaither, Associate Director of the Texas Advanced Computing Center

� Visualization for a range of applications including computational fluid dynamics, bioinformatics and geophysical modeling

� Complex rendering and multiprocessing of graphics and images for researchers in a diverse range of fields

� Sun Fire V880z and a Scalable Distributed Sun Fire 6800 visualization system

Data

ComputeandVisualization

GridSystemSoftware

Sun Grid Reference ArchitecturesProven Solutions, Updated Regularly

Implementing a Grid Solution� Compute Grid Infrastructure Solution

Sun Fire V20z based Compute Grid Rack System Opteron

Sun Fire E25K Supercluster 144 SPARC cores/node

Customer Ready Systems factory built to specification

StarterCluster for Bioinformatics low cost entry point

� Grid Reference Architecture Designed, tested, tuned, benchmarked and documented

Opteron with Gbit Ethernet, Myrinet and Infiniband

Scalable NFS using clustered QFS

Sun BluePrint: Implementing Globus on Solaris

� Grid Professional Services Worldwide Practice Workshop, architecture assessment, design and implementation

Case Study: CellTech,Life Sciences� Better

Up to 99% Utilization Heterogeneous resources

� Faster 6-10x improvement in productivity

� Cheaper Lower cost of software licenses

Sun enabled us to meet our peak demandSun enabled us to meet our peak demand

turnaround requirements and run turnaround requirements and run

algorithms that simply weren't possible beforealgorithms that simply weren't possible before

Alan Hart, Celltech Alan Hart, Celltech

Sun's Grid Ecosystem

Grid System Software

Grid Data Center

Global Grid Practice

Alliances

Data Compute Visual

Adrian Cockcroft

[email protected] Everywhere.

www.sun.com/solutions/hpc/

Grid EverywhereSun's HPTC/Grid Strategy

grid everywhere - jpgrid.org · maestro: nasa jet propulsion laboratory first steps towards the...

Documents