Providing HPC Resources forProviding HPC Resources forPhilips Research and PartnersPhilips Research and Partners
Ronald van DrielRonald van Driel
Senior TechnologistSenior Technologist
Research ICTResearch ICT
April 23April 23rdrd, 2007, 2007
2Research
Agenda
• Introduction Philips and Research
• Research Interest in HPC
• HPC and Grid Related Activities
• Vision on Future Directions
3Research
Royal Philips Electronics
• One of the largest global electronics company with sales in– 2006 of EUR 26,976 million– Q4 2006 of EUR 8,128 million
• Founded in 1891
• Multinational workforce of 121,700 employees (January 2007)
• Active in the areas of Healthcare, Lifestyle and Technology
• Manufacturing sites in 28 countries, sales outlets in 150 countries
• R&D expenditures EUR 1,619 mln (2006)
Headquarters: Amsterdam, The Netherlands
4Research
Significant new product introductions in 2006/2007
Medical Systems Lighting
Consumer Electronics
DAP
EP Navigator
Gemini Time-of-Flight
PET/CT
SureSigns VM
BrightView SPECT
Aluminum Juicer
Wardrobe care
Williams F1 shaver
Wake up lightVOIP
phone
Portable Media Devices
Ambient Experience
HD Category
ActiLume
Edore
Living Colours
Luxeon ® K2LED
Cosmopolis
6Research
Regional Representation
Briarcliff
Redhill
Aachen
Shanghai
Eindhoven
Bangalore
Hamburg
Staffed by around 1,800 people
9Research
World-class technology centre of high tech companies working together in the development of new technologies
910,000 sq.m Over 50 nationalities 7,000-8,000 people by 2008 more than EUR 500 million
investment by Philips Comprises a broad set of world-
class facilities and expertise, that are shared with partners:• Micro-and Nanotechnology
infrastructure and services (MiPlaza), HD Studio, EMC3, LSF, etc.
10Research10
High Tech Campus EindhovenStrength through diversity & networks
Corporate innovators
Consultancy & services
Research institutes
Economic development companies
11Research
Global Research ICT Organization
CIOPatrick Stemkens
Adaptive & DedicatedICT
(ADICT)
GlobalInfrastructures
BusinessInformation
Systems
InfrastructureDevelopment
InfrastructureManagement
InformationServices
12Research
Agenda
• Introduction Philips and Research
• Research Interest in HPC
• HPC and Grid Related Activities
• Vision on Future Directions
13Research
History of Central Compute Facilities at Philips Research Eindhoven
1980 1985 1990 1995 2000 2005
IBM Mainframe – VM/CMS & MVS
VAX/VMS
Apollo Workstations
GRID
Central Compute Cluster (NXA)HP-UX Linux
Silicon Graphics – Video/Audio/Graphics
14Research
Observations I
• Developments in application fieldse.g. Medical Imaging and Bioinformatics
– Many sensors and high-resolution sensors: data explosion– Correlation of data from different sources (e.g. SPECT/CT)
• Open Innovation– Need to cooperate to get resources and/or knowledge in the
right place at the right time: “Support the technology chain”
• Infrastructure– Cope with the “data explosion” and “large scale data analysis”– Large computations, but much can be parallelized or distributed– Must facilitate secure collaboration with partners
From model driven to data driven experiments
15Research
Observations II
Cannot be done by a single entityMust share and cooperate
Big growth in application complexity. Resource needs of some application fields may grow orders of magnitudeWill lead to huge growth in data storage and compute capacity needs.
And today's business innovation is done in an Open Innovation setting: Collaboration in virtual teams is the way of working.(Industry, Universities & Institutes) Needs to be supported by applications and IT infrastructures.
16Research
IT Resources Importance I
Storage:The ability to cope with very large and often distributed datasets.
Visualization: "Virtual reality" techniques are becoming an increasingly important part of applications or workflows.
17Research
IT Resources Importance II
Compute: A substantial increase in demand for computer power and the ability to run data intensive tasks.
Networks:Enabling technology for providing fast access to distributed resources.
18Research
Research Interest in HPC?
• A number of program, project and application fields at Philips Research will require (or may benefit from) high performance computing:– Medical imaging to improve image quality and accuracy.– Molecular Medicine and Bioinformatics will be faced with an
explosive growth of data.– Large simulations and inverse problems
• Lighting armature optics for solid state lighting
– System in Package (SIP) and Solid State Lighting may need to introduce 3D models and simulations.
• HPC capabilities might yield attractive added value for potential partners of Philips Research.– Important to be part of the ECO system
19Research
Agenda
• Introduction Philips and Research
• Research Interest in HPC
• HPC and Grid Related Activities
• Vision on Future Directions
20Research
Research ICT HPC and Grid Activities
• Philips Research ICT is trying to close the gap between “demanding” jobs that researchers have and the supporting IT infrastructures.
• Translate upcoming IT technology in new strategic options• Developed various HPC/grid demonstrator aimed at
helping researchers to do their calculations faster…• Installed a cluster that is connect to DutchGrid to
understand and investigate the technology– Access via “gLite” or “UNICORE” grid middleware software– Allows for more demanding, jobs to be sent to other (possibly
external) “clusters”
• Participate in “Virtual Laboratory for e-Science” (VL-e) and “BIG-GRID” projects
21Research
Example: Researchers Workflow
Short cycles require maximum flexibility: PC/WS or local cluster
Secure access to large data and compute resources: shared on a site. Debug possibilitiesUsually executed on a “Site Cluster”
Full simulations using specialized data, compute and visualization infrastructures.
• E.g. medical imaging, bioinformatics simulation and search, 3D multi-physics modeling, system in package simulation, etc.
• Many projects do not follow the complete flow, but stay at one or more stages.
Application dependent solutions e.g. dedicated computers or a bunch of chips
Workflow View
Experimenting with algorithmor numerical set-up
Validation and regression tests
Pre-production tests withfull implementation
Production implementation
22Research
Example: Researchers Workflow
Short cycles require maximum flexibility: PC/WS or local cluster
Secure access to large data and compute resources: shared on a site. Debug possibilitiesThis “Site Cluster” is part of “DutchGrid”!
Full simulations using specialized data, compute and visualization infrastructures.
• E.g. medical imaging, bioinformatics simulation and search, 3D multi-physics modeling, system in package simulation, etc.
• Many projects do not follow the complete flow, but stay at one or more stages.
Application dependent solutions e.g. dedicated computers or a bunch of chips
Workflow View
Experimenting with algorithmor numerical set-up
Validation and regression tests
Pre-production tests withfull implementation
Production implementation
Site (small) cluster at HTC Eindhoven Part of VL-e infrastructure Infrastructure provider for VL-e Maximum control and fast feedback Priority setting and security can be guaranteed
23Research
Grid Demonstrator – Aachen / Eindhoven
• PET System Simulation• Monte Carlo simulation of the
scanner up to the detection of scintillation light– ~32 CPU days on 2.8 GHz Xeon– 100 x 1 GB output
• System-level simulation of the detector electronics (Mona/Lisa)– 2 phase post processing
• Simulation can be highly parallelized
24Research
PET Simulation Workflow Schematic
motronI2 O2
motronI.. O..
motronI1 O1
I
MotronS
IN S
Mona
MCA C2
MCA C..
MCA C1
1 GB
80 MB
100 MB
< 1 KB
25Research
Cluster/Server/WorkStationCluster/Server/WorkStation
Meta-Scheduler / Scheduler
OpenPBS, SGE, LSF, Condor,Others (SDK)
GridAgent
GridXpert Synergy : Architecture
GridManager(J2EE) PKI RDBMS
GridXpert Synergy:
• GridManager– Full J2EE standards– Embedded Certificate
Authority / PKI for all resources and users
– API available• Enterprise portal (EIP)• Import/export XML
modelization
• GridAgent– Job Management– 3rd party data transfer
between GridAgent (GridFTP)
– Based on Globus development
– Integration with lead schedulers
Web Browser
Jdbc
HTTPS / RSL Globus
HTTPS / HTML Soap
ApplicationXML
CLI
Meta-Scheduler / Scheduler
OpenPBS, SGE, LSF, Condor,Others (SDK)
GridAgent
26Research
Philips jobs at NIKHEF
SPECT Simulation
• 20,000 jobs of ~3.2 CPU hours each• Executed at NIKHEF and SARA clusters• Access via LCG-2 Grid middleware software• Single Photon Emission Computed Tomography
27Research
Grid Portal for PET System SimulationPositron Emission Tomography• Application developed by
Researcher in Aachen.• Design (graphical/web)
portals– Core services too complex
to present to scientists
• Based on GridSphere web portal technology (JSR-168)
• Initial version developed with assistance from GridwiseTech
28Research
Simulate Full HDTV Data Processing Chain
• Picture quality enhancement• Picture rate-up conversion• Pixel plus …
Input OutputAlgorithm 1 Algorithm n
~0.5 TB ~1500 CPU hours ~2.2 TB
Original (Frame repetition)
Measured > 50 MB/s data transfer rates between Amsterdam and Eindhoven over GigaPort connection
29Research
GAMA ResearchHealthcare Systems Architecture research group
WindowsIDL, Pride
LinuxGlobus
GatewayClient
LAN LANDutchGridResources
The GAMA architecture: computational Grids• Architecture for solving compute-intensive medical applications • Minimally invasive: Running on Grid as alternative, easy fall back to local versions• Simultaneously provides different sets of services to multiple users and applications• Adaptive to various healthcare applications• Example: Brain imaging, the High Angular Fiber tracking (HAFT) application
Clinical trials of HAFT software at AMC are being executed on VL-e cluster at HTC in Eindhoven.
30Research
VL-e Cluster HTC EindhovenSuper
ComputeCenters
BIG GRID,Sara,
Jülich, etceteras
Fas
t In
tern
et
Bioinformaticsinfrastructure
Internet
DutchGrid
Secure Environment extern Philips
Compute Cluster
Cluster Load Balancing
Grid Access LayerExternal
Oracle
Database
Computeservers
31Research
UNICOREUniform Interface to Computing Resources• Speedup Monte Carlo based
simulation package developed on Microsoft Windows platform
• Ported application to Linux• Allows for easy parallelization
using loosely coupled jobs• Deploy UNICORN client on
Microsoft Windows platform• Python script to create XML file
that is loaded in OpenMolGRID• Project defined to developed
application specific UNICORE plug-in
32Research
Agenda
• Introduction Philips and Research
• Research Interest in HPC
• HPC and Grid Related Activities
• Vision on Future Directions
33Research
Vision: Shared Resources and InfrastructureEnd-user view
• Applications “flow through infrastructure” and can be run and used anywhere:
Boosts sharing and collaboration
• Users decide whether to keep data in internal secure environment or to put in shared space
• Applications might be split in secure and non-secure parts
34Research
Vision: Shared Resources and Infrastructure Infrastructure view
• Grid is next generation shared infrastructure comprising distributed data storage capacity, (visualization) equipment and compute clusters:– Capitalize on our extensive experiences with
centralized compute and data clusters– Other equipment will shift to the shared infrastructure
too (e.g. VL-e)
35Research
Vision: Shared Resources and Infrastructure Success factors
• End-user guidance and support is essential!!• Deploy workbenches for application fields• Continuous development, be a leader• Standardisation of middleware layer
– e.g. gLite, UNICORE and Globus
• QoS and AAA are key in e-Science production infrastructures
• Basic building blocks are data storage and (low latency) compute clusters