EMEA HPTC Virtual Team *Other brands and names are the property of their respective owners © Copyright 2002-2003 Intel Corporation High-Performance Computing.

EMEA HPTC Virtual Team

*Other brands and names are the property of their respective owners© Copyright 2002-2003 Intel Corporation

High-Performance Computing on Intel® Architecture

EUROGRID Overview

Klaus-Dieter Oertel

Intel GmbH, Solutions & Software Group

[email protected]

2nd EUROGRID Workshop @ CGW03

Cracow, Oct 29, 2003




Outline

EUROGRID and Intel EUROGRID Goals Technical Basis: UNICORE EUROGRID Results




Pallas HPC group has developed the UNICORE Client and has participated in EUROGRID

Intel has acquired the Pallas HPC group (Sep 12)

Pallas HPC will form the Cologne Software Lab (CGSL) within the Software Solutions Group (SSG)

CGSL’s Grid Mission:Make Intel a leading player in Grid technology

CGSL takes Pallas role in EUROGRID

EUROGRID and Intel




EUROGRID Vision

Build a European Grid infrastructure

that gives users

a seamless, secure access to

High Performance Computing resources

and that advances computational science in

Europe




EUROGRID Goals European HPC centers setup GRID testbeds

– Seamless access to HPC resources, Multi-site jobs

Interfaces for important applications– Bio-/Chemical, Meteo, CAE

Industrial applications and GRID usage scenarios– ASP services for engineers– Intranet solutions for distributed organizations

Development of UNICORE extensions– E.g. resource broker, efficient data transfer, . . .

Contribute to Global GRID development




EUROGRID Partners

HPC Centers FZ Jülich (D, Admin. Coord.) CSCS Manno (CH) ICM Warsaw (PL) IDRIS Paris (F) Univ Bergen (N) Univ Manchester (UK)

Users DWD (D) EADS (F) T-Systems (D) (Assistant Partner)

Duration: Nov 2000 – Jan 2004Volume: 33 person years, 3.5 MEuro cost

funded by European Commission Grant No. IST–1999–20247

Integration Pallas (D) (Project Coordinator)

FLE (UK) (Assistant Partner)




EUROGRID Geography




The Technical Basis: UNICORE

Vertical Grid System for (HPC) Compute Grids

Graphical client

Sophisticated security concept

Plugins for existing applications(Nastran, Fluent, CPMD, ...)

Object-oriented architecture, implemented in Java




UNICORE Client

Graphical Interface to UNICORE Grids Open Source under Pallas Community License Job Preparation, Monitoring and Control Complex Workflows File Management Certificate Handling Plug-in Interface




UNICORE Job Model

Directed acyclic graph of– tasks (computational or data

transfer)– sub–jobs (to be executed at

another system)– temporal dependencies

Attach resource requeststo tasks– capacity (CPU time, disk, …)– capability (MPI–2, profiling, …)– can do static tests of jobs– can do resource brokering …

If Success Execute LM Model

Import Model Data

LM–JobT3E–1200

Process–LMZampano

Postprocess

Export Data

Notify user

Transfer Data




UNICORE Network Job Supervisor

UNICORE Architecture

Client can run anywhere Gateway as single point of

entry NJS incarnates and

schedules jobs UDB (User Database)

contains user login information

IDB (Incarnation Database) contains resource information and job incarnation rules

TSI (Target System Interface) interfaces to local batch system

Cray NQE PBSPro

TSI

Network Job Supervisor

Gateway

Client(s)

Insecure Internet

IDB

TSIIDB

UDB




UNICORE – Security Model User authentication

– permanent user certificate– stored in protected keystore on the user PC/workstation– certificate passed downstream as part of the job

Job authentication– job description is signed with the user’s private key– downstream sites can check whether job is authentic– no trust required between UNICORE sites

Software authentication– client and gateway use certificates for mutual authentication– NJS can use a special “user certificate” for authentication to secondary

sites or take the role of a client User authorization

– performed by the UNICORE server according to local procedures (mostly based on local userid)




UNICORE Plugins – Extend the User Interface

UNICORE client has a plugin interface – Create GUIs that support important applications– Plugin simplifies data entry for application– Plugin can support application-specific resources– Plugin constructs (complicated) workflows automatically

Plugin will use UNICORE Client to– Submit the application job– Monitor and control the application job

Helpful features– End-users concentrate on applications– Extended consistency checks

Existing Plugins– CPMD molecular dynamics code– Star-CD, Fluent, Nastran, …




Example: Gaussian98 Plugin




Results of EUROGRID

Setup and operation of European GridsBio-GRID, Meteo-GRID, HPC-GRID

Industrial use of Grids startedCAE-GRID: EADS (Airbus), T-Systems

Grid-enabling of many applications via plugin interface

Development and integration of new technology Alternative File Transfer, Resource Broker, Billing Services, Cost Estimation, Application Coupling, Interactive Access




Bio-GRID

biogrid.icm.edu.pl

GRID for biomolecular simulations

Interfaces to existing biological andchemical codes




Bio-GRID: Applications

ADF: IDRIS CPMD: FZJ Gamess: ICM, FZJ, UoM, Parallab Gaussian98: ICM, FZJ, UoM, Parallab, IDRIS Molcas: FZJ, Parallab Molpro: Parallab Turbomole: FZJ Amber: ICM, FZJ, UoM, Parallab, IDRIS CHARMM: ICM, Parallab Gromos96: ICM




Bio-GRID: Plugins

Gaussian98 Amber 6.0 PDB search plugin

– Search mirrors of the PDB (Protein Data Bank) database and choose database providing fastest transfer

– Visualize results using Jmol, JMV (Java Molecular Viewer), RasMol/RasWin, or user defined application

List Jobs Plugin– Checks all available Vsites– Gather information about user jobs and their status

File Filter Plugin– Access files in job directory of running jobs




Meteo-GRID

Develop a relocatable version of DWD‘s weather prediction model

Goal:‘Weather prediction-on-demand‘ as an ASP solution




CAE-GRID

Coupled simulations of aircrafts(e.g. structure and electromagnetics)

Goal:internal HPC portal for EADS engineers




CAE-GRID

CAE-GRID at EADS is not publicly available EADS has developed a plug-in for acoustic &

electromagnetic applications– Combining 4 applications in one job




HPC-GRID Demonstration of a European HPC GRID testbed Sharing of competence and know-how Agreement on certification, access policies, ...

CRAY T3E 900 (32 PE)NEC SX4B/2ALinux Cluster (4 PE)

SGI SGI O2000 (128 PE)

CRAY T3E - 1200 (816 PE)FUJITSU VPP300 (8 PE)SGI O2000 (128 PE)SGI O3000 (256 PE)

IBM SP3 (8 PE)NEC SX5 cluster (40 PE)IBM Power4 (256 PE, 1.3 TFLOPS)COMPAQ Linux Cluster (24 PE)

Linux Intel Cluster (36 PE)CRAY T3E - 600 (512 PE)CRAY T3E - 1200 (512 PE)

FZJ




Alternative File Transfer

Scenario: multi-site job where a huge amount of data has to be transferred between sites

UNICORE internal transfer mechanism is based on “slow” Java streams

UNICORE server has been extended to use different file transfer mechanisms e.g. rcp

Use AFTGridFTP in EUROGRID which is: GridFTP as an alternative file transfer mode




AFTGridFTP

GridFTP:– Developed in the Globus project– FTP (file transfer protocol) with new features for grid applications

such as parallel and partial file transfer– Requires Globus proxy certificates for operation

AFTGridFTP is transparently used for transfers if– The job contains a Transfer Task– The sites are set-up for AFTGridFTP– The Client has loaded the proxy certificate plugin

• Generates a proxy certificate based on the user certificate• The proxy certificate is provided to GridFTP in the job directory• Proxy certificate plugin has been developed in the GRIP (Grid

Interoperability) project




AFTGridFTP

Performance results depend on– File size– Form of the file: compressed or uncompressed– Number of parallel streams

AFTGridFTP can be 15 times faster than default UNICORE Java streams




Resource Broker

User specifies job with resource requests (e.g. a certain application resource)

Send job to Broker site which might resolve to a hierarchy of Broker sites

Broker site replies with a list of offers where to run the job

Each offer is a ticket which provides a time-limited service guarantee

Accepting one offer includes the ticket in the job Submit the job to the selected site Valid ticket in the job indicates the acceptance of the

QoS offer




Resource Broker




Resource Broker




ASP Extensions - Billing Services

Software accounting system on the server– Generate accounting data

(userid, used software, hostname, cpu time, …)– Record the accounting data in a database– Independent of UNICORE

Billing plugin– Retrieve the accounting data and corresponding

prices– Create and provide the billing data to the user




ASP Extensions - Billing Services




ASP Extensions – Cost Estimation

Generic cost estimation is not possible for CAE applications in general

The LS-Dyna application supports cost estimation LS-Dyna plugin has been extended for cost

estimation– A pre-job starts the LS-Dyna simulation for some calculation

cycles– The necessary simulation time is estimated based on

extrapolation– Hardware and software costs are calculated and displayed

in the plugin




ASP Extensions – Cost Estimation




Application Coupling

A plug-in has been developed which supports the execution of CORBA coupled applications under UNICORE– Query the status of a CORBA Name Server– Or remove the context references– Start the coupled applications

The applications exchange data independently from UNICORE using only the CORBA framework




Application Coupling




Interactive Access

An Interactive Access plugin has been developed for– Interactive execution of pre-defined short

commands– Interactive steering of batch jobs

The job directory can be inspected at run time The user can reconnect to running interactive

batch jobs




Interactive Access




More information

EUROGRID: www.eurogrid.org

UNICORE: www.unicore.org

Pallas / Intel: www.pallas.com

GRIP: www.grid-interoperability.org




Questions?

EMEA HPTC Virtual Team *Other brands and names are the property of their respective owners © Copyright 2002-2003 Intel Corporation High-Performance Computing.

Documents

eurogrid intel

intel slide

eurogrid eurogrid

emea hptc virtual team

respective owners copyright

unicore eurogrid results

nd eurogrid workshop

hpc resources