Globus Grid Middleware Toolkit Otto Sievert CSE 225 8 June, 2000

Preview:

Citation preview

Globus

Grid Middleware Toolkit

Otto Sievert

CSE 225

8 June, 2000

Globe

and other European Grid Activities

Otto Sievert

CSE 225

8 June, 2000

EGRID

• European Grid Community

• Collaborative community, not a standards group

• Commercial and Academic Interests

• www.egrid.org

European Tour

• Netherlands• Germany• Poland• Italy• Sweden

Grid Theme 1

• Be very (very) careful when choosing a project name.

Amsterdam: Globe

• Vrije Universiteit– Maarten van Steen– Andrew Tanenbaum

• “Middleware to facilitate

large-scale distributed applications”– Web focus– object-based coherency

Globe Uniqueness

• Too much data, too little resources (bandwidth, etc.)

• Caching helps

• Data Coherency integral to Cache Policy

• Release constraint of a single replication/ distribution policy for all objects– example: web pages

IMAGES

COUNTS

HTML

Globe Object

• Physically Distributed• Replicated

• Distribution Policy

Globe Local Object

• 4 subobjects (minimum)

• Modularity

IMAGES

SEMANTICS

REPLICATION

CONTROL

COMMUNICATION

Globe Binding

1. Name server

2. Object handle

3. Location service

4. Contact points

5. Choose point

6. Repository

7. Protocol

8. Bind!

NAMINGSERVICE

LOCATIONSERVICE

IMPLEMENTATIONREPOSITORY

CLIENTPROCESS

1

2

3

4

8

6

7

DISTRIBUTEDSHAREDOBJECT

Legion Binding

• Two-stage name resolution

• Binding agent• No implementation

repository

BINDINGAGENT

CLIENTOBJECT

1

2

3

4

5

SERVEROBJECT

Autopilot Binding

1. Sensor registers with the sole manager

2. AP client requests sensors from the manager

3. Manager returns available sensors

4. Client and sensor communicate directly

MANAGERCLIENT

1

2

3

4 SENSOR

Globe Claim: Remote Object Model Lacks Replication

• Globe– objects can be

replicated

– still maintain one state

– allows complex coherency policies

• Legion– in theory, supports

replication

– replicated state

– allows some but not all coherency policies

– in practice is not allowed

NAMINGSERVICE

LOCATIONSERVICE

IMPLEMENTATIONREPOSITORY

CLIENTPROCESS

1

2

3

4

8

6

7

DISTRIBUTEDSHAREDOBJECT

Globe Architecture

• Why all these servers?– separate naming from

locating– allow extensible

binding protocols (?)

Grid Theme 2: Communication

Centralized– NWS

– Globus MDS*

• Simple Management• Single Point of Failure

Distributed– NetSolve

– Fran’s Sensor Net

• Complex Management

• Scalable

Mixed– Legion

– Globe

– Autopilot

*

***

*** * *

*

*

***

*

Globe Implementation Example

• Set of HTML/image/ Java files

• One semantics subobject• browsers aren’t that

extensible, so …use gateway

SEMANTICS

REPLICATION

CONTROL

COMMUNICATION

browser

gateway

http://globe.foo.edu:8989/dir/file.html

Globe Example (cont’d)

• Replication Policies– Object-based

• “Permanent store”

• “Object-initiated store”

– Client-based• “Client-initiated store”

• How is this any better than what we have now?

Globe Live Demo ...

Globe Location Service

• Scalability Questions

• Tree Heirarchy– again Legion-like in its worst-case behavior

• Typical sol’n assumes mobile client

• Globe sol’n assumes mobile software

Does This Work?

• Single experiment - 5 wk. web server trace

• compare– no caching– various complex replication/coherency policies– automatic adaptive policy

• Results– (essentially) any replication scheme wins big– individual object adaptivity didn’t perform well

Globe: Conclusion

• Explicit coherency is clearly a Good Thing

• Security?

• Representative implementation?

Germany: Cactus

• Albert Einstein Institute, Potsdam– Thomas Radke

– Ed Seidel

• Distributed Astrophysics• Software Engineering• NCSA “hot code”

Cactus

• Separate CS from disciplinary science– Flesh = CS architecture

– Thorns = science modules

• 2-stage compilation– encapsulation

– modularity

– reuse

Cactus Compilation

• Two stage– Permanently bind thorns [Perl]– Compile binary [C++/F77]

• Efficient– Don’t carry unneeded thorn info

Grid Theme 3: Applications

• Numerical, or Non– Computation vs. Specialization

• Performance Measures:– Execution Time– Scale– Efficiency– Distribution

Grid Theme 4: Transparency

• Ease of Use vs. High Performance– As system becomes opaque, EoU increases– As system becomes opaque, Perf decreases– Where is the balance?

Germany: Gridware

• 1999 San Jose-based merger of two companies: Genias GmbH and Chord (U.S.)

• CoDINE– Compute farm load balancing system– Recently adopted by Sun™

• PaTENT– WinNT MPI

Grid Theme 5: Commoditization

• Reuse is strong in the Grid– Resources (Beowulf)– Middleware (Globus, PaTENT)– Applications (Cactus)

• Industry is influential– Largest grid apps in use today are commercial– Grid-ernet is profitable

To This Point ...

Resource

Middleware X X

Application X X

Commercial X

Germany: Unicore

• UNIform Computer Resources (German SC access)• Goal is to provide uniform access to high performance

computers - painful to learn– OS details

– Data storage conventions

– Administration policies

• 3 phase project:I self-contained jobs

II remote data access

III concurrent remote execution

Unicore (cont’d)

• How is this done?– Web (Java) user interface– X.509 authentication– Network Job Supervisor

• interprets Abstract Job Objects

• manages jobs and data

• interfaces with local batch systems (like LoadLeveler and CoDINE)

• vs. Globus?

Poland: POL-34

• National Grid• Very like the system used

by Unicore, a collection of widely-distributed parallel computers

• Tree-connected ATM network

POL-34

• Yellow = 2 Mb/s• Red = 34 Mb/s• Cyan = 155 Mb/s• Single administrative

domain via Virtual Users (skirting the grid issue)

• Use Load Sharing Facility (LSF)

Italy: SARA

• University of Lecce, Italy Giovanni Aloisio (with Roy Williams of Caltech)

• Synthetic Aperature Radar Atlas– Distributed data-intensive app

– Alan Su and the UCSD AppLeS group is involved

SARA Architecture

• The goal: easy, fast, efficient retrieval and processing of SAR image data

• Issues– data is distributed, stored

in tracks

– complex hierarchical system

• Prototypical grid app

Sweden: Computational Steering

• Parallelldatorcentrum Royal Institute of Technology Per Oster

• Using Globus and the Visualization Toolkit (VTK) to steer a single CFD code.

• Little data available• Eclipsed by Autopilot

Grid Theme 6: Heterogeneity

• Some attempt to hide it– Globus, CORBA, Java/Jini

• Some take advantage of it– Netsolve, Ninf

• Some characterize and manage it– AppLeS (SARA), NWS

To This Point ...

Resource X

Middleware X X X X

Application X X X X

Commercial X

Conclusion

• Explored Globe, Cactus, and other EuroGrid favorites in the context of– Communication architectures– Grid application characteristics– Grid transparency– Commodity computing influence– Grid heterogeneity

Network Weather Service (NWS)

• Rich Wolski, U. Tenn.• Monitors and Predicts

Grid Resources– network latency,

bandwidth

– CPU load, avail. Mem.

• Central NWS data server

• nws.npaci.edu/NWS

Recommended