HPC: a Paradigm for the ANSYS Software
Transcript of HPC: a Paradigm for the ANSYS Software
HPC: a Paradigm for HPC: a Paradigm for the ANSYS Softwarethe ANSYS Software
Michel RochetteMichel RochetteR&D DirectorR&D Director
© 2009 ANSYS, Inc. All rights reserved. 1 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 1 ANSYS, Inc. Proprietary
Outline
• ANSYS, Inc ,
• High Performance Computing
• Performance - Fluids
• Performance - StructurePerformance Structure
• Healthcare Challenges and HPC
• Cerebral Aneurysms
• Osteoporosis• Osteoporosis
• Conclusion
© 2009 ANSYS, Inc. All rights reserved. 2 ANSYS, Inc. Proprietary
World’s LargestSimulation CommunitySimulation Community
20 3Network of sales channel partners in 40+ countriesApprox. 1,700 employees / 40+ sales offices on 3 continents
© 2009 ANSYS, Inc. All rights reserved. 3 ANSYS, Inc. Proprietary
20 major development centers on 3 continents
© 2008 ANSYS, Inc. All rights reserved. 2/08 after Q4 2007 ANSYS, Inc. Proprietary
2008 & 2009 Target % of revenue spending on R&D: 15%
World’s LargestSimulation CommunitySimulation Community
15 000 T t l C t> 15,000 Total Customers> 250,000 Commercial Seats> 220,000 University Seats> 200 Channel Partners> 150 I d t P t
20 3Network of sales channel partners in 40+ countriesApprox. 1,800 employees / 40+ sales offices on 3 continents
> 150 Industry Partners
© 2009 ANSYS, Inc. All rights reserved. 4 ANSYS, Inc. Proprietary
20 major development centers on 3 continents
© 2008 ANSYS, Inc. All rights reserved. 2/08 after Q4 2007 ANSYS, Inc. Proprietary
2008 & 2009 Target % of revenue spending on R&D: 15%
In-Depth TechnologySpanning Multiple DomainsSpanning Multiple Domains
Technical Breadth
Tet/PrismHex/Hex Core
StructuredUnstructured
Multi-ZoneBody-Fitted Cartesian
Patch IndependentMore…
Meshing
Conduction
ThermalCompressible
FluidsTech Quasi static (Low Freq)
ElectromagneticsStructuralLarge Displacements
y
ConvectionRadiationPhase ChangeMass Transport
IncompressibleLaminar Flow
phnical D
ep
Full WaveJoule Heating
Finite StrainContactMultibody DynamicsShock & Vibration
TurbulenceMultiphase Flow
ppth
Eddy currentsCurrent flow
Implicit & Explicit
More…
pReacting FlowMore…
St d St t T i t H i & M d l
Circuit CouplingMore…More…
© 2009 ANSYS, Inc. All rights reserved. 5 ANSYS, Inc. Proprietary
Steady-State, Transient, Harmonic & Modal
Linear & Nonlinear
High Performance Computing
High Performance ComputingComputingComputing
© 2009 ANSYS, Inc. All rights reserved. 6 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 6 ANSYS, Inc. Proprietary
HPC Value Proposition
• High Performance Computing (HPC) -and ANSYS parallel performance -provides you with:– Faster turnaround timeFaster turnaround time
• For single simulations, or multiple simultaneous jobs
Ability to consider bigger more detailed– Ability to consider bigger, more-detailed models• Full assemblies, direct CAD-to-Mesh
– Ability to consider more complex physics• Time varying simulations, multiphysics
– Capacity to consider multiple design p y p goptions• Parametric modeling, DOE
© 2009 ANSYS, Inc. All rights reserved. 7 ANSYS, Inc. Proprietary
Higher productivity, improved insight, and faster development of better products
Parallel Performance for Higher-Fidelity Simulation Higher Fidelity Simulation
• More complete models– From single blade passage to full 360– Computational workload increases by 90xp y– Turnaround time maintained at ~7 hours
• ANSYS CFX parallel 90 process
• More complex physics• More complex physics– From steady “RANS” model to unsteady “LES”– Computational workload increases by >100x– File I/O time also increases significantly– Turnaround time maintained at ~8 hours
• ANSYS FLUENT Parallel 256 processANSYS FLUENT Parallel 256 process
© 2009 ANSYS, Inc. All rights reserved. 8 ANSYS, Inc. Proprietary
ANSYS 12.0 – HPC Leadership
• ANSYS 12 0 development included a major investment inANSYS 12.0 development included a major investment in HPC:
• Focus on processor and platform optimization– Optimized for the latest multicore chips– Support for clusters running Linux and Windows HPC Server 2008
• Outstanding parallel scaling improvementsOutsta d g pa a e sca g p o e e ts• Linear performance out to 1000’s of cores (Fluids)• Teraflop performance and scaling to 512 core (Mechanical)
S t f ll l I/O• Support for parallel I/O• Support for “billion” scale simulations
“ANSYS has consistently focused software design on the full range of computing platforms ― from desktop to supercomputer ― with their technology yielding great performance on the latest high-performance computing (HPC) solutions. This ensures that our mutual customers can tackle ever more complex and high-fidelity simulations while still achieving
© 2009 ANSYS, Inc. All rights reserved. 9 ANSYS, Inc. Proprietary
p g y gthe turnaround time required for product development decision making.”― Richard Dracott, General Manager, HPC Division, Intel, U.S.A.
Performance - FluidsPerformance - Fluids
© 2009 ANSYS, Inc. All rights reserved. 10 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 10 ANSYS, Inc. Proprietary
ANSYS CFD 12.0Quad-Core Processor ScalingQuad-Core Processor Scaling
• ANSYS CFD 12 0ANSYS FLUENT and ANSYS CFX
Dual Socket Quad Core Performance• ANSYS CFD 12.0 incorporates extensive tuning for multicore 8 00
8,00
7,16 7 18 7,58 7 55
Dual-Socket Quad-Core PerformanceIntel Xeon 5500 (“Nehalem”)
gperformance– Compiler optimizations,
optimized core mapping and 5,00
6,00
7,00
8,00 , 7,18
6,23
7,55
6,796,47
Spee
dup
optimized core mapping and binding, enhanced cache re-use
• Result: ~80-95% of ideal 2,00
3,00
4,00
,
Cor
e So
lver
S
speedup!
8 Way Parallel Value 12
48
0,00
1,00
8-Way Parallel ValueRun 3-5M cell CFD simulationsReduce execution time by 6-7x
Answers in 2 hours instead of 12
1
© 2009 ANSYS, Inc. All rights reserved. 11 ANSYS, Inc. Proprietary
Answers in 2 hours, instead of 12
ANSYS CFD 12.0Multi-Processor ScalingMulti-Processor Scaling
CFX Solver scalabilityANSYS CFX 12.0 Scaling PerformanceAMD O t 2218
80 00
100.00
120.00AMD Opteron 2218
Truck_Poly_14M Benchmark
40.00
60.00
80.00
Spee
dup
Ideal
Transonic Airfoil 10M Nodes
0.00
20.00
0 16 32 48 64 80 96
b f
Infiniband
GigabitEth t
• ANSYS FLUENT 12.0 delivers dramatic scaling P ll l S li V l
number of coresNumber of Cores
improvement in this release• ANSYS CFX 12.0 scaling also outstanding
Parallel Scaling Value14M cell simulation, 512-way
Assume 500 iterationsAns ers in 1 ho r instead of 150
© 2009 ANSYS, Inc. All rights reserved. 12 ANSYS, Inc. Proprietary
Answers in 1 hour, instead of 150
ANSYS CFD 12.0Massive Simulation SupportMassive Simulation Support
• ANSYS FLUENT 12.0 includes architecture updates to support today’s largest simulations
• Demonstrated ability to solve model sizes yof one billion cells and above.
• Sample timing: 3.5 hours for 500 iterations on 768 cores; 1.3 hours for
Solver Speedup
2048
ANSYS FLUENT 12.0 Scale-OutSGI Altix ICE 8200EX – 2048 Cores
;data write.
• Demonstrated linear scaling at 1024 cores(111M cell benchmark)
991
1581
1024
1536
er S
peed
up Truck_111M Benchmark
( )• 78% of ideal at 2048 cores
128265
521
0
512
Solv
e
Speedup
Ideal
Mega-Modeling ValueConsider highest fidelity models.Develop detailed understanding.
© 2009 ANSYS, Inc. All rights reserved. 13 ANSYS, Inc. Proprietary
00 512 1024 1536 2048
CoresNumber of Cores
p gGet answers back 1000x faster.
ANSYS CFD 12.0Parallel and Serial - I/OParallel – and Serial - I/O
• Parallel I/O provides unique scaling ANSYS FLUENT 12.0breakthrough at large scale – Addresses the I/O bottleneck– New file pdat
Serial I/O Parallel I/O
ANSYS FLUENT 12.0
New file .pdat– Uses MPI-IO and parallel file system– All nodes write to a single file– Files are interoperable
• Traditional “serial” I/O also improved• ~2x speedup in read/write ANSYS CFX
– Parallelized data compression• ~3-9x speedup for data write ANSYS FLUENT
Parallel I/O ValueParallel I/O Value10M cell simulation, 192-way
1000 iterations, 100 data writesI/O reduced to 20% from 70% of time
© 2009 ANSYS, Inc. All rights reserved. 14 ANSYS, Inc. Proprietary
I/O reduced to 20% from 70% of timeAnswers in 1 hour, instead of 2
Performance - StructuresPerformance - Structures
© 2009 ANSYS, Inc. All rights reserved. 15 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 15 ANSYS, Inc. Proprietary
ANSYS Mechanical 12.0Quad-Core Processor ScalingQuad-Core Processor Scaling
• ANSYS Mechanical 12 0 ANSYS Mechanical 12.0• ANSYS Mechanical 12.0 delivers improved scaling on latest quad-core processors from AMD and Intel 8,83
Dual-Socket Quad-Core PerformanceIntel Xeon 5500 Processor ("Nehalem")
from AMD and Intel• Fully loaded cores
• Typically 5 - 6x speedup on 6,00
7,00
8,00
9,008,00
5,745,46
8,83
5,53 5,83
peed
upyp y p pdual-socket quad-core systems
• 3 hour simulation now completed in 30 minutes! 2 00
3,00
4,00
5,00
,
4,55
Cor
e So
lver
Sp
completed in 30 minutes!
12
480,00
1,00
2,00
8-Way Parallel Value1
Data from Cray CX-1 Personal Supercomputer
3-5M DOF mechanical simulationsReduce execution time by 5-6x
Answers in minutes, instead of hours
© 2009 ANSYS, Inc. All rights reserved. 16 ANSYS, Inc. Proprietary
Data from Cray CX 1 Personal Supercomputer Running Microsoft Windows HPC Server 2008On your desktop, or on a cluster
ANSYS Mechanical 12.0Multi-processor ScalingMulti-processor Scaling
ANSYS Mechanical 12.010M DOF Distributed ANSYS PCG Solver
ANSYS Mechanical 12.03M DOF Distributed ANSYS Sparse Solver
50607080
eedu
p
10M DOF Distributed ANSYS PCG SolverIntel Xeon 5500 Processor Series ("Nehalem")
3000
4000
5000
AMD Opteron 2360 ("Barcelona")QLogic TrueScale Infiniband
Tim
e (s
ec)
1020304050
Cor
e So
lver
Spe
0
1000
2000
3000
Cor
e So
lver
T
01 2 4 8 32 64 128
C
Number of Cores
16 32 64 128 256Number of Cores
• ANSYS Mechanical 12.0 delivers enhanced S S ec a ca 0 de e s e a cedscaling due to:
• improved domain decomposition• Improved load balancing
Parallel Scaling Value3M DOF benchmark
~12 hours (one processor)1 h (16 )
p g• distributed matrix generation
• Teraflop performance demonstrated on 512 cores
~1 hour (16 cores)~15 minutes (128 cores)
Value of 12 hours saved: $1200
© 2009 ANSYS, Inc. All rights reserved. 17 ANSYS, Inc. Proprietary
Healthcare Challenges Healthcare Challenges
© 2009 ANSYS, Inc. All rights reserved. 18 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 18 ANSYS, Inc. Proprietary
The Healthcare Worldwide ChallengeChallenge
• Healthcare expenses in North Americap– In 2007, healthcare spending in the U.S. reached $2.3 trillion. It is
projected to reach $4.2 trillion by 2016.I 2005 th U S t 16% f it d ti d t (GDP)– In 2005, the U.S. spent 16% of its gross domestic product (GDP) on healthcare.
• Worldwide healthcare concernsWorldwide healthcare concerns– Healthcare spending accounted for:
• 10.9% of GDP in Switzerland• 10.7% in Germany• 9.7% in Canada• 9.5% in France (Organization for Economic Cooperation and Development)% ( g p p )
• Increasing or maintaining the quality and comfort of healthcare in the future will be a major challenge.
© 2009 ANSYS, Inc. All rights reserved. 19 ANSYS, Inc. Proprietary
Engineering Challenges
• Cost-effectiveness– Prevention and early detection– Treatment efficacy– Personalized medicine– New treatment approval
Fl d h t t f i th
• Comfort and social trends– Portable medicine
Flow and heat transfer in the human eye
– Less invasive surgery – Minimizing impact on active life
• Innovation– Drug efficacy
© 2009 ANSYS, Inc. All rights reserved. 20 ANSYS, Inc. Proprietary
g y– VPH (virtual physiological human) Contours of oxygen concentration plotted on
vertical planes inside incubator
Courtesy Silesian University of Technology.
Maturity of Engineering Simulation TechnologySimulation Technology
• Existing medical imaging techniques to provide accurate g g g q ppatient specific morphology (0.1 mm)
• The newest techniques (ultrasound elastography, MRI) will id di l t d l iti ( bl tprovide displacements and velocities (comparable to our
simulation results)• Robustness and speed for fluid flow and structural analysisRobustness and speed for fluid flow and structural analysis• Major progress in CAD import and meshing
– Automatic process flow from medical imaging to clinician p g ginterpretation of simulation results
• Access to living tissues material properties • Valuable clinical demonstrations using ANSYS software suite
© 2009 ANSYS, Inc. All rights reserved. 21 ANSYS, Inc. Proprietary
ObjectivesObjectives
• To use existing simulation software on patient g pspecific organs, bones, tissues, arteries …in order to “predict” a specific biomedical behavior p p
• This simulation, based on the combination of bio-imaging and simulation technologies will help theimaging and simulation technologies, will help the clinicians for diagnosis or surgery planning
© 2009 ANSYS, Inc. All rights reserved. 22 ANSYS, Inc. Proprietary
ApplicationsApplications• Osteoporosis: risk of fracture, cement injection
• Cerebral Aneurysms: risk of rupture stent• Cerebral Aneurysms: risk of rupture, stent
• Coronary, Carotid: Modulography, rupture of atherome plaque, morphology, diagnostics, surgery planningp gy, g , g y p g
• Per-Op surgery software for Abdominal Aortic Aneurysms (from 2D angiography to 3D image of the deformed artery)
• Ophthalmology: Cornea and crystalline lens stiffness, Patient specific presbyopia laser surgery
• Early Cancer Detection using inverse problem between medical imaging (MRI or US) and simulation
F 2D X i t f ll 3D d l• From 2D X-ray views to a full 3D model
• From 4D images to organ motion and morphing (cardiac simulation)
© 2009 ANSYS, Inc. All rights reserved. 23 ANSYS, Inc. Proprietary
• Aesthetic Surgery (soft tissues deformation)
• Dental implant, …
HPC challenges for Medical SimulationSimulation
• Patient specific simulation for diagnosis and surgeryp g g y– Collaborative Development with Labs and Clinicians to develop the vertical
application – Integration of medical images and engineering simulation for patient specificIntegration of medical images and engineering simulation for patient specific
clinical applications– Multiscale simulation: protein, cell, tissue, organ, body, population
• Coupling between lower and upper scales• Coupling between lower and upper scales
– Accurate engineering simulation on patient data to be computed in a reasonable time
• Some applications require quasi real time (per op surgery)• Some applications require quasi real time (per-op surgery)
– Database of simulation results taking into account the human variability• Huge design of experiments
N t ti f di l i i i i l ti• Next generation of medical imaging using simulation technologies– 4D images (US, MRI) will be filtered using model motion and morphing
© 2009 ANSYS, Inc. All rights reserved. 24 ANSYS, Inc. Proprietary
g ( , ) g p g– Incomplete Medical Simulation models could be completed using imaging
data (Boundary Conditions, materials)
HPC Focus on 2 medical applications
HPC Focus on 2 medical applicationsapplicationsapplications
Cerebral AneurysmCerebral Aneurysm
© 2009 ANSYS, Inc. All rights reserved. 25 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 25 ANSYS, Inc. Proprietary
Cerebrovascular Diseases
• Cerebrovascular disease: StrokeCerebrovascular disease: Stroke– 15% hemorrhagic (aneurysms)– 85% ischemic (stenosis )85% ischemic (stenosis …)
• Aneurysm prevalence is between 2% and 4%.In US 12 millions people have aneurysm– In US 12 millions people have aneurysm
• The incidence (risk of rupture) is 10/100,000 per yearThis risk co ld be pre ented sing endo asc lar• This risk could be prevented using endovascular surgery (stent)
This surgery is expensive and risky (2 to 5%)– This surgery is expensive and risky (2 to 5%)
© 2009 ANSYS, Inc. All rights reserved. 26 ANSYS, Inc. Proprietary
Status Today
Initiation
GrowthR t
10/100K/yearRupture
© 2009 ANSYS, Inc. All rights reserved. 27 ANSYS, Inc. Proprietary
2-4%2,000 - 4,000 / 100KSource: Humphrey JD, Cardiovascular
solid mechanics 2001
Risk Assessment: State of the Art
• Today a cerebral aneurysm is considered as a virtualToday a cerebral aneurysm is considered as a virtual time bomb in your head
• Hemodynamics (blood flow) governs the pathology• Hemodynamics (blood flow) governs the pathology from the initiation to the growth and rupture
Aneurysm risk assessment only based on morphology• Aneurysm risk assessment only based on morphology
• The decision for surgery is based on simple criteria b i d h li i l i f iabout aneurysm size and other clinical information– E.g. Cerebral Aneurysms beyond 5mm are treated;
• Even if an existing cerebral aneurysm has a statistical risk around 1% it is a wise decision to treat it through
© 2009 ANSYS, Inc. All rights reserved. 28 ANSYS, Inc. Proprietary
endovascular surgery
@neurIST - www.aneurist.org
• @neurIST is a major multidisciplinary European initiative within the SixthEuropean initiative within the Sixth Framework Programme*
• The project brings togetherThe project brings together– Neurosurgeons– Neuroradiologists– Epidemiologists– Engineers– Biologists– Computer scientists from 32 European institutions
Members of the consortiumMembers of the consortium
• The aim of the project– To develop a usable interface for personalised
risk assessment and treatment of patients with pcerebral aneurysm and subarachnoid haemorrhage
• 4 years project Jan 2006 until Dec 2009
© 2009 ANSYS, Inc. All rights reserved. 29 ANSYS, Inc. Proprietary
4 years project, Jan 2006 until Dec 2009
Cerebral Aneurysm
• Medical imaging software can extract Vessel surface extraction
g gpatient-specific aneurysm geometry and blood flow
• ANSYS suite of software creates a patient specific virtual model of thepatient-specific virtual model of the aneurysm region
• Risk of rupture evaluatedWall shear stress in aneurysm
Patient medical image (3DRA)
– Wall shear stress in aneurysm during the cardiac cycle
• Proactive treatments can be prescribed
Reduction to region of interest
• HPC requirements for unstented and stented aneurysms
© 2009 ANSYS, Inc. All rights reserved. 30 ANSYS, Inc. Proprietary
Computational meshHaemodynamic results
@neurIST Haemodynamic Toolchain
OsteoporosisVPHOP
OsteoporosisVPHOPVPHOPVPHOP
© 2009 ANSYS, Inc. All rights reserved. 31 ANSYS, Inc. Proprietary© 2009 ANSYS, Inc. All rights reserved. 31 ANSYS, Inc. Proprietary
Osteoporosis is a killer
• OP fractures kill as many women as breast ycancer.
• 30 to 50% of all women and 15 to 30% of all• 30 to 50% of all women and 15 to 30% of all men will face an osteoporotic fracture in their lifetimelifetime.
• 4,000,000 fractures every year cost Europe €30,000,000,000.
• Forcast to double by 2050.Forcast to double by 2050.• 250,000 elders will dies of related
li ti ithi 12 th ll th ill© 2009 ANSYS, Inc. All rights reserved. 32 ANSYS, Inc. Proprietary
complications within 12 month; all others will remain impaired.
Not enough technology
• The technology in current clinicalThe technology in current clinical practice is clearly insufficient.
• The accuracy in predicting• The accuracy in predicting fractures is as low as 60%.E if th d t• Even if we see the drugs are not working we wait for the fracture,
d l th i ll fi itand only then surgically fix it.
© 2009 ANSYS, Inc. All rights reserved. 33 ANSYS, Inc. Proprietary
Better chances to deliver
• Bone physiology is as complex as any other organ
• But the biomechanics of bone fracture is in itself a purely ymechanical event.
• This is one of the few domainsThis is one of the few domains where organ-level models already achieve predictivealready achieve predictive accuracies of over 90%
© 2009 ANSYS, Inc. All rights reserved. 34 ANSYS, Inc. Proprietary
Predictive, Personalised
• P2 medicine: the VPHOP project will make possible Predictive and Personalised (P2) medicine for osteoporosis– Predictive: multiscale models representing the
skeletal mechanobiology from the whole body down to the molecular constituents, simulate the skeletalthe molecular constituents, simulate the skeletal loading in various conditions and predict if the bones will fracture in each of them
– Personalised: The multiscale model is personalised using specific patient information. The more available information, the more personalised the modelinformation, the more personalised the model becomes
© 2009 ANSYS, Inc. All rights reserved. 35 ANSYS, Inc. Proprietary
The Hypermodel: primary systemic relationshipsprimary systemic relationships
Tissue-level
Boundary Conditions
ConstitutiveEquation
Organ-level model
model
BoR
emod
Body-level model
onedelling
FailureCriterion
Cell-levelmodel
Constituent-levelmodel
Criterion
© 2009 ANSYS, Inc. All rights reserved. 36 ANSYS, Inc. Proprietary
Organ level
Resorption ReversalHYPERMODELHYPERMODEL
Resorption Reversal
Resting Formation
© 2009 ANSYS, Inc. All rights reserved. 37 ANSYS, Inc. Proprietary
HPC Challenges for Clinical PracticePractice
• Multiscale Simulation• Multiscale Simulation– Computation of the risk of femoral or vertebral fracture– Simulation taking into account the microstructure and cell level g
remodelling– Several hundreds millions of cells on a huge cluster (5000 cores
in CINECA cluster)in CINECA cluster)
• Interventional treatment planning– Predicting the most clinical location within each bone– Predicting the most clinical location within each bone – Predicting the changes in risk due to interventional augmentation– HPC requirements for the compatibility of clinical practice
© 2009 ANSYS, Inc. All rights reserved. 38 ANSYS, Inc. Proprietary
Conclusion
• HPC is strategic for Computer Aided Engineering:g p g g– To consider more detailed models (size x100)– To consider more complex physics p p y– To consider parametric modeling and design of
experiments• Extreme scalability demonstrated on thousands of
cores for CFD• Extreme scalability for structural is the current step• These HPC innovations will open new applications in
personalized medicine• A new era for health technologies
© 2009 ANSYS, Inc. All rights reserved. 39 ANSYS, Inc. Proprietary