High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High...
Transcript of High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High...
![Page 1: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/1.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing forFast Hybrid Simulations
Boris Jeremic and Guanzhou Jie
Department of Civil and Environmental EngineeringUniversity of California, Davis
CU–NEES 2006 FHT Workshop, Nov. 2-3 2006
![Page 2: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/2.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Outline
![Page 3: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/3.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
Outline
![Page 4: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/4.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
Need for HPC: Hybrid Simulations
Real Time Simulations
I Structural and Geotechnical component testing is ratedependent
I Algorithmic measures are only partially fixing the problem
I Goal: use real time computational simulations for hybridtesting
I In other words, use real time computational simulations tocontrol the test
![Page 5: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/5.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Fast Computer Simulations: Various Models andAlgorithms
I Implicit vs Explicit (on local and global level, both offeradvantages and disadvantages)
I Hierarchy of material models (use only required level ofsophistication for the purpose)
I Reduced Order Modeling (ROM) now available fornonlinear models
![Page 6: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/6.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Standard Application Programing Interface: OpenSeesI Core framework (class libraries UCB, modifications by
UCD),I Material models (UCD, UCB, UCSD, UW, Stanford ...)I Finite Elements (UCD, UCB, Stanford ...)I Procedures (loading, dynamics, solution control) (UCD,
UCB, Stanford ...)I Parallel Simulations (Plastic Domain Decomposition, UCD)
I Official Manuals available (UCB, personally do not like it)I Our own effort UCD, covering only our efforts, but will
expand to cover other areas of interest.I TheoryI Command (interactive)I Examples (verification and validation)
![Page 7: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/7.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Fast Computer Simulations: Programming (CompSci)
I Mix and match efficient programming languages (C++, C,Fortran, Tcl, Python...)
I Efficient compilers (GNU, Portland Group, Intel (KAI),Metroworks)
I HPC programming techniques (loop unrolling, templatemetaprograming...)
![Page 8: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/8.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Fast Computer Simulations: Hardware
I Fast sequential machines (balance multiple levels of cacheand RAM, disk, CPUs clock speed, data bus speed...)
I Parallel machines (SMPs and DMPs, networking...)
I It it time for another Finite Element Machine (NASAendeavor from some 20 years ago)
![Page 9: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/9.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Symmetric Multiprocessor Parallel (SMP) Machines
I aka Shared Memory Parallel machines (dual coremachines count too)
I Shared memory communications, bandwidth 555.831MB/s, latency: 7.897micro seconds
I Cache coherence problem
I Not scalable above 16 or so CPUs
I High price
I Fine grained parallelism (matrix–vector multiply...)
I Automatic tools for parallelization (somewhat effective)
![Page 10: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/10.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Distributed Memory Parallel (DMP) Machines
I aka Beowulf clusters (or clusters of SUNs as available atCU in ’95)
I Networked communications: bandwidth 74.935 MB/s,latency: 62.487 micro seconds
I Message passing paradigm for parallel programming
I Scalable to a large number of machines (CPUs,thousands)
I Very good price/performance ratio (low)
I Communication controls performance
I problems arise with power consumption, cooling andplacement area
![Page 11: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/11.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Clusters of Clusters
I aka DMPs of SMPs
I SMPs can be used as DMPs, with a very fast network
I Cluster a number of SMPs (dual core machines count too)
I Even better price/performance ratio (low)
I Multiple levels of communication performance
I Very efficient and flexible setup!
I Example: GeoWulf (8× 2 + 1× 4 + 1× 8 + 4× 1 andadding as funding permits)
I Example: Hemisphere (64× 2)
![Page 12: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/12.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Computing (HPC) for FHT
High Performance Software and Hardware
Interdependence of Software and HardwareI Specifics of the problem dictate the hardware designI Example #1: large scale (20,000+ DoF) problems (SFSI for
example, solids for soils, beams and plates/shells forstructure) can be efficiently simulated on DMPs, whileSMPs might be too restrictive (size)
I Example #2: Smaller scale problems (1,000+ DoF) can beefficiently simulated on SMPs and/or DMPs (mostly in realtime)
I Example #3: Small scale problems (below 1,000 DoF) arebest suited for SMPs can be efficiently simulated on SMPsand even DMPs (in real time)
I PDD is designed to take advantage of all the abovearchitectures automatically
![Page 13: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/13.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
Outline
![Page 14: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/14.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
Efficient parallel Simulations for Inelastic Models
PDD Method: Design Goals
I Graph partitioning→ balance multiple phasessimultaneously, while also minimizing the inter-processorcommunications costs
I It is a multi-objective optimization problem (minimize boththe inter-processor communications, the data redistributioncosts and create balanced partitions)
I Take into the account (deterministic or probabilistic):I heterogeneous element loads that change in each iterationI heterogeneous processor performance (multiple
generations nodes)I inter-processor communications (LAN or WAN)I data redistribution costs
![Page 15: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/15.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
Efficient parallel Simulations for Inelastic Models
PDD Method: Implementation
I Perform global optimization for both (a) internal statedetermination and system of equations solution phases
I Adaptive partitioning done using ParMETIS
I Parallel direct and iterative system of equations solverswithin PETSC interface. Iterative: CG, GMRES;preconditioners BLOCKSOLVE95, HYPRE and SPAI;Direct: SPOOLES, MUMPS, SuperLU_DIST
I OpenSees: standard interface and framework
I Works on SMPs, local DMPs, grids of computers
![Page 16: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/16.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
PDD: Current Features
Features
I Initial domain partitioning
I Adaptive domain repartitioning depending on CPUimbalance, LAN and/or WAN performance
I Repartitioning works with loads, constraints..., allnecessary movable objects
I Available for elements (solid, structural) with standardOpenSees API (sendSelf, RecvSelf, timer or CL weightestimate)
I Scalable to a large number of CPUs
I Performance tuning (GeoWulf (UCD), TeachingMachine(UCD), Hemisphere (CU Boulder), IA64 (SDSC))
![Page 17: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/17.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
PDD: Current Features
Absolute Speedup Overview
0 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 1280
8
16
24
32
40
48
56
64
72
80
88
96
104
112
120
128
Number of CPUs
Par
alle
l Spe
edup
Theoretical S
peedup
398,721 DOFs
118,275 DOFs
68,451 DOFs
32,091 DOFs
17,604 DOFs
![Page 18: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/18.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Plastic Domain Decomposition Method
PDD: Current Features
PDD vs DD Speedup Overview
![Page 19: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/19.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
Outline
![Page 20: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/20.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
SFSI Model
Detailed 3D FEM modelI Construction processI Two types of soil: stiff soil (UT, UCD), soft soil (Bay Mud)I Deconvolution of given surface ground motionsI Use of the DRM (Bielak et al.) for seismic inputI Piles→ beam-column elements in soil holesI Structural model developed at UCB (Fenves et al.)I Element size issues (filtering of frequencies)
model size el. size fcutoff min. G/Gmax γ
12K 1.0 m 10 Hz 1.0 <0.5 %15K 0.9 m >3 Hz 0.08 1.0 %150K 0.3 m 10 Hz 0.08 1.0 %500K 0.15 m 10 Hz 0.04 3.0 %
![Page 21: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/21.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
SFSI Model
SFSI FEM Mesh (one of)
Eight cases considered for sand (S) and/or clay (C) beneatheach bent.
![Page 22: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/22.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
SFSI Model
Soil and Structure Displacement Response
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0 10 20 30 40 50 60
−0.2
−0.1
0
0.1
0.2
0.3
Time (s)
Dis
plac
emen
t (m
)
SSS CCC SSC CCS SCS CSC SCC CSS
![Page 23: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/23.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
SFSI Model
Soil and Structure Acceleration Spectra
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.05
0.1
0.15
0.2
0.25
0.3
Period (s)
Four
ier A
mpl
itude
(m/s
2 )
SSS CCC SSC CCS SCS CSC SCC CSS Freefield Input Motion
![Page 24: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/24.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
SFSI Model
Top of Columns Moment Response (Redistribution)
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
0 10 20 30 40 50 60−4000
−3000
−2000
−1000
0
1000
2000
3000
4000
Time (s)
Mom
ent (
kN*m
)
SSS CCC SSC CCS SCS CSC SCC CSS
![Page 25: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/25.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
Fast Hybrid Testing Model
I Real time (fast) simulation is imperative for Fast HybridTesting (FHT)!
I Hierarchy of models with different levels of sophisticationI 2D frameI 3D framesI 2D excitationsI 3D excitationsI material modelingI SFSI!
![Page 26: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/26.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
Fast Hybrid Testing Computational Setup
I Parallel computer Hemisphere
I Computer nodes used as DMP of SMPs (cluster of cluster)
I Extremely fast, dedicated interconnect (fiber optics) fromHemisphere to FHT lab (latency below 0.1ms)
I Allows for possible expansion of computational resources(at the FHT lab), while maintaining LAN with Hemisphere.
I This will create a cluster of clusters, a very efficient(price/performance) design.
I Current Hemisphere users can (transparently) useadditional computational resources as well
![Page 27: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/27.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
Fast Hybrid Testing (outside) Networking
I Connection from UCD (GeoWulf) is not fast enough
I Physics (speed of light) issue,
I GeoWulf (UCD) - 1428.76km - FHT;
I Surface fiber optics, one way trip ∼ 7.0ms or ideally invacuum ∼ 4.8ms,
I In reality it is 10× that much (and it is probabilistic)
I Similarly SDSC (NEESit) is 1329.79km away from FHT
![Page 28: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/28.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
SAC Frame Model, PDD and FHT
I Basic model is a 2D frame on fixed foundations
I Elements and material models: Elastic beams and plastichinge elements
I Sophisticated model (realistic) is a 3D frame with SFSIeffects, 3D seismic wave propagation
I Choice of # of computational nodes on Hemisphereparallel machine depends on a the size and sophisticationof the model
![Page 29: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/29.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
Computational Trials
I Very small model (2D, or 3D but no SFSI) uses efficientlyup to 8 CPUs
I More sophisticated model (3D, SFSI) uses efficiently 32 ormore CPUs
I Very sophisticated model (3D struct., SFSI, 3D seismicwaves) uses efficiently all 128 CPUs (and would use moreif available)
I For a very sophisticated model the challenge isperformance (real time, fast)
![Page 30: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/30.jpg)
HPC for FHT: Plastic Domain Decomposition Method
High Performance Simulation Examples
FHT model
Parallel Computer Hemisphere Performance
0 4 8 12 16 20 24 28 320
4
8
12
16
20
24
28
32
Number of CPUs
Para
llel S
peed
up
Theoretical32,212 DOFs
![Page 31: High Performance Computing for Fast Hybrid …sokocalo.engr.ucdavis.edu/~jeremic/ · High Performance Computing for ... Department of Civil and Environmental Engineering University](https://reader031.fdocuments.net/reader031/viewer/2022021901/5b81724b7f8b9a2b678c6173/html5/thumbnails/31.jpg)
HPC for FHT: Plastic Domain Decomposition Method
Summary
Summary
I High performance, parallel simulations are neededI Fast Hybrid Testing (control)I High fidelity model based simulationsI Soil–Foundation–Structure control during seismic eventsI . . .
I PDD parallel finite element code is GPL (available)
I http://sokocalo.engr.ucdavis.edu/~jeremic/PDD/