gpu-based high-performance computing for urban · pdf filenonlinear cyclic pushover analysis...
TRANSCRIPT
Xinzheng LU
Institute of Disaster Prevention and Mitigation
Department of Civil Engineering, Tsinghua University
Beijing 100084, P.R. China
GPU-based high-performance
computing for urban seismic
damage prediction and visualization
The Sixth Kwang-Hua Forum, Shanghai, China, 2014
Outlines
Introduction
Program Framework
Performance Benchmark
Case Study
Realistic Visualization
Conclusions
Introduction
China is subjected to most serious earthquake disaster
threats in the world
Earthquake occurs in cities will cause tremendous
casualties and damage
Scientific prediction of urban seismic damage is an
important task
Beichuan City, 2008 Tangshan City, 1976
Introduction
Methods for urban seismic damage simulation
Based on probability matrices
ATC-13
Based on capacity curve and
response spectrum
HAZUS, AEBM
Problems:
• SDOF model
• Pushover analysis
• Demand Spectra
• ….
Introduction
Single structure Urban region
• Detailed structural information
• One building
• Limited structural information
• Hundreds of thousands of buildings
“Nonlinear Time History
Analysis of a City!”
Introduction
University of Tokyo
Integrated Earthquake Simulation
Supercomputer with
traditional CPU platform
Expensive
Complex
High maintenance costs
Introduction
GPU (Graphic Processing Unit)
Video cardGPU
Comparison between the CPU and the GPU
CPU: 2~8 cores
GPU: hundreds/thousands
of cores
GPU-powered THA of Single Bld.
Fiber beam element + Multi-layer shell element
Collapse simulation of Z15 in Beijing (H=550m)
Collapse simulation of Shanghai Tower (H=630m)
Collapse simulation of reinforced concrete high-rise
building induced by extreme earthquakes, Earthquake
Engineering & Structural Dynamics, 2013, 42(5)
Collapse simulation of a super high-rise building
subjected to extremely strong earthquakes. Science
China Technological Sciences, 2011, 54(10)
GPU-powered THA of Single Bld.
Platform Hardware Price Solver
CPU Intel Core i7-3970X 3.5GHz (Fastest CPU in the market)
US$2406 SparseSYM of
OpenSees
GPU Intel Core i7-4770X 3.4GHz &NVIDIA Geforce GTX Titan
US$2307 CuSPSolver of
OpenSees
409
27.5
0
100
200
300
400
500
CPU SparseSYM GPU CuSPSolver
Com
putin
g tim
e (
h)
-0.8
-0.4
0
0.4
0.8
1.2
0 5 10 15 20 25 30 35 40 45
To
p d
isp
./m
Time/s
PGA=400gal
CPU GPU
0
20
40
60
80
100
120
140
0 0.2 0.4 0.6 0.8
Sto
ry
Displecement envelope/m
PGA=400gal
CPU GPU
14.87x faster!
The advantages for using GPU
GPU-powered THA of Urban
Region
Seismic computing for
normal buildings Computing features of GPU
Simple model, few degree-of
freedom in a single building
Relatively weak performance
of a single core
No interaction between
buildings Fewer data exchange
A huge number of buildings Suitable for parallel computing
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Computational
Model
Inter-story
Hysteretic
Model
Building
Performance
Database
Seismic Damage
Simulation Based on
GPU/CPU Coop.
Computing
Program Framework
1. Building Models
2. Building Performance Database
3. Parallel Computing Method
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Building Models
Computational Model
Multi-story concentrated-mass shear (MCS) model
Moderate workload
Consider higher-order vibration modes & velocity pulses
Damage locations on different stories can be obtained
Suitable for GPU computing
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Building Models
Inter-story hysteretic model
Backbone curve
Trilinear, 5 parameters
Hysteretic model
Modified-Clough
Bilinear elasto-plastic
Pinching
V
Δ
k0
ηk0
Vy
βVy
O Δc
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Building Performance Database
For Regular Buildings
Based on the HAZUS performance database
Parameter Set Selection
According to building macro-parameters
Structural types, Numbers of stories, Construction Period
19 building types proposed in HAZUS are adopted
HAZUS Parameters MCS model Parameters
convert
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
For Special Buildings:
Building Performance Database
Objective Special
Building
Refined Nurmical
(RN) Model
Multi-story
Concentrated-mass
Shear (MCS)
Model
Nonlinear Cyclic Pushover Analysis
Normalized by the weight of one story
Nonlinear Cyclic
Pushover Curve
of RN model
(Each Story)
Nonlinear Cyclic Pushover Analysis
Nonlinear Cyclic
Pushover Curve
of MCS model
(Each story)
Coincide or Not
Adjust the Parameters
No
Parameters of inter-story
backbone curve and
hysteretic model
Yes
Building Performance Database
Validation (six-story RC frame)
Refined FE model Inter-story hysteretic model
Top displacement Inter-story drift
Parallel Computing Method
Pre-analysis Module
Seismic Analysis Module
Post-analysis Module
CPU
GPU
CPUInput Data for nonlinear
THA
Copy Data to Graphic
Memory
Assign Tasks for GPU
THA for Building #2THA for Building #1 THA for Building #3
Copy Data to Host Memory
Output Damage &
Response Data
...THA for Each Building
Performance Benchmark
CPU/GPU cooperative vs. CPU only
1,024 buildings, numbers of stories and structural
types are random generated
Earthquake record: El Centro, 40 s, PGA: 200 cm/s2
Time of data input and output is not included
Platforms Hardware Compliers
CPU Intel Core i3 530 @2.93GHz &
DDR3 4G 1333MHz. Microsoft Visual C++ 2008 SP1
GPU/CPU
cooperative
Intel Celeron E3200 @ 2.4GHz &
NVIDIA GeForce GTX 460 1GB.
Microsoft Visual C++ 2008 SP1 &
CUDA 4.2
Performance Benchmark
Weak-scaling benchmarks for the two platforms
39x speedup when computing 1024 buildings
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
0
50
100
150
200
250
300
0 256 512 768 1024
Co
mp
uti
ng
Tim
e (s
)
Number of Buildings
253.65 s
7.06 s
276.62 s
11.84 s
CPU Single
CPU Double
GPU/CPU Coop. Single
GPU/CPU Coop. Double
Case Study
A medium-sized urban area in China
4,225 buildings
Global view
Damage on
different stories
Case Study
Desktop Computer
4,255 buildings
40 s time-history analysis
Accomplished in 216 s
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Local view
Damage on
different stories
Case Study
Desktop Computer
4,255 buildings
40 s time-history analysis
Accomplished in 216 s
A coarse-grained parallel approach for seismic damage simulations of urban areas based on refined models and GPU/CPU cooperative computing, Advances in Engineering
Software, 2014
Case Study
Local view
Peak acceleration
on different stories
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm/s
2)
Near-Field with Pulses
None Slight Moderate Extensive Complete
Case Study
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm2/s
)
Far-Field
None Slight Moderate Extensive Complete
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm2/s
)
Far-Field
None Slight Moderate Extensive Complete
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm2/s
)
Near-Field without Pulses
None Slight Moderate Extensive Complete
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm2/s
)
Near-Field without Pulses
None Slight Moderate Extensive Complete
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm
2/s
)
Near-Field with Pulses
None Slight Moderate Extensive Complete
0 500 1000 1500 2000 2500 3000 3500 4000 4500
70
200
400
Number of Buildings
PG
A (
cm
2/s
)
Near-Field with Pulses
None Slight Moderate Extensive Complete
Far-field
Near-field
without
pulses
Near-field
with pulses
MCS Model SDOF Model
MCS model: velocity pulses can be considered
Number of Buildings Number of Buildings
Visualization problem
Realistic visualization
Rescue and transportation planning
MCS model cannot simulate process of building collapse.
MCS model
(Criterion of collapse)
Real earthquake disaster
(Include building collapse)
Building collapse
Physics Engine Solutions
Physics engine
A computer program for real-time dynamic calculation, good
at multi-body dynamics.
Widely used in computer graphics, video games and film.
Example of physics engine
Seismic damage simulation in urban areas based on a high-fidelity structural model and a physics engine, Natural Hazards, 2014
Physics Engine Solutions
Finite element results
Finite element results + physics engine based-debris
Physics engine-driven visualization of deactivated elements and its application in bridge collapse simulation, Automation in Construction, 2013
Collapse simulation
The process of collapse simulation in physics engine
Integrate MCS model and physics engine
Seismic damage simulation in urban areas based on a high-fidelity structural model and a physics engine, Natural Hazards, 2014
Collapse simulation
30 frame/s
High-efficient collapse simulation.
Application for Tsinghua Campus
The campus of
Tsinghua University
Conclusions
GPU-based high-performance
computing for urban seismic
damage prediction and
visualization
Fast Cheap High-fidelity Realistic
With Texture Disp. Contour Damage State
Thank you for your attention!
Nonlinear time history analysis:
from mega-structures to cities?