2. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
History of Adventure_Thermal
1996 Massively Parallel Computing for Large Scale Finite Elements R.Shioya
1997-2002 Adventure Project
2000.12.01 AdvSolid-0.8b
2000 領域分割法による 3 次元熱伝導解析 石川格
2001.11.14 AdvThermal-0.2b
2002 大規模並列弱連成解析 今村栄太
2002.03.01 AdvThermal-0.3b, AdvSolid-1.0(BDD)
2003 Study of Speed-up Technique of Large Scale Parallel FE. A.Mukaddes
2003.09.16 AdvSolid-1.1
2004 領域分割型有限要素法による超並列計算 荻野正雄
2005.07.11 AdvThermal-1.0(BDD)
3. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Functions added to ADVENTURE_Thermal
Version: 0.6b
01. Configure options ( install and compile)
02. BDD and BDD-DIAG solver
03. Bug fixed advthermal-p and advthermal-s
Version: 1.0
01. “mkbc4th” tool for convection and radiation boundary
conditions
02. “makefem_thermal” tool for flux boundary conditions
03. English manual ( with some example problems and tool
explanation)
04. Japanese manual (with some example problems, tool
explanation and BDD explanation )
Next Version
01. IBDD and IBDD-DIAG method
02. Time dependent boundary conditions
7. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
vectorNormal:n
][retemperatuExternal:
tcoefficienferHeat trans:
generationheatInternal:
heatSpecifice:
Density:
][onappliedeTemperatur:u
onappliedfluxHeat:Q
retemperatuofderivativeTime:
t
u
tyconductiviThermal:
rflux vectoHeat:q
][eTemperatur:
C
u
Q
Cu
f
C
C
Cu
o
C
o
o
α
ρ
λ
Γ
Γ
∂
∂
u
_
Q
on
on0
on0)(
indiv
ingrad
Γ=
Γ=−⋅
Γ=−−⋅
Ω+−=
∂
∂
Ω−=
uu
Qnq
uunq
fq
t
u
c
uq
CCC
α
ρ
λ
uΓ
Heat conductive problem
8. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Finite Element (quadratic tetrahedral) Approximation
)}({}{}{)}({
][})]({[})]({[
)}({}})]{({[
}{
})]({[
tRRRtf
KuKuK
tfuuK
dt
ud
uC
hfq
hc
++=
+=
=+
convectionsurfacefromarisingVector:)}({
generationheatinternalfromarisingVector:}{
fluxheatfromarisingVector:}{
convectiontorelatedMatrix:][
conductiontorelatedMatrix:})]({[
matrixStifness:})]({[
matrixeCapacitanc:}]({[
retemperatunodalofVector:}{
tR
R
R
K
uK
uK
uC
u
h
f
q
h
c
9. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Time Discretized Equation
nnnnnnn
uKtCftuKtC }){][)1(]([}{}){][]([ 1
∆−−+∆=∆+ +
ϕϕ
sec)increment(time:t
size(mm)meshminimum:
3
)schemedifferenceBackward(1
∆
=
=
h
m
ϕ
This equation is solved by Domain Decomposition Method in
each time integration loop
)2)(1(
2
)1( 2
++
≤
∆
−
mmh
t
ϕλ
Stability condition :
10. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Domain decomposition
Interface
Domain Decomposition Method (DDM).
gSuB =
fKu =
Original finite element problem:
Analysis domain
After eliminating interior DOF, we get a reduced
problem on interface, called the interface
problem,
local Schur
complement
( ) ( ) ( ) ( )( ) ( )i
IB
i
II
Ti
IB
i
BB
i KKKKS
1−
−=
Schur complement matrix( ) ( ) ( )∑= =
N
i
Tiii RSRS 1
( )
( )
( ) ( )iTi
i
i
:R
:f
:K
Ω→Ω restriction operator
local stiffness matrix
local RHS vector
I : interior dofs, B : interface dofs
where
Interior
17. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
hddmsrc/hddm_types.h
typedef struct {
int domid ; /* subdomain id in its own part */
int gdomid ;/* global id of domid */
int partid ;
int nel ; /* number of elements */
int nnd ; /* nodes */
int ntemp ; /* temp b.c. */
int nflux ; /* flux b.c. */
int nconv ; /* conv b.c. */
int nradi ; /* radi b.c. */
int ninbd ; /* inner-domain b.c. */
int nsharenode;
int nd_elm ;
int nd_elm2 ;
int *nop ;
int *ndindex ;
Inbc *inbd ;
Bcond *bctemp ;
Bcond *bcflux ;
Bconv *bcconv ;
Bradi *bcradi ;
Sharenode *sharenode;
Slist *bclist ; /*+ Set of
boundary conditions +*/
char elm_type[128] ;
double *temp ;
double *oldtmp ;
double *force ;
double *flux ;
double *crd ;
/* material */
int have_matid ;
int *matid ;
int nmat ;
Material *mat ;
} DomMesh ;
18. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
typedef struct {
int partid ;
int n_part ;
int n_domain ; /* Number of in domains */
int n_outdom ; /* Number of out domains */
int t_nnd ; /* Total number of nodes */
int t_infree ; /* Total inner freedom */
int t_outfree ; /* Total outer freedom */
int t_midfree ; /* Total middler freedom */
int sum_nel ; /* Sumation of nel of all domains */
int sum_nnd ; /* Sumation of nnd of all domains */
int sum_nel_orig, sum_nnd_orig ;
int sum_ntemp ; /* Sumation of the number of templacement B.C. of all domains */
int sum_nflux ; /* Sumation of the number of flux B.C. of all domains */
int sum_nconv ; /* Sumation of the number of conv B.C. of all domains */
int sum_nradi ; /* Sumation of the number of radi B.C. of all domains */
int sum_ninbd ; /* Sumation of the number of inner boundary nodes */
double *crd ;
int have_matid ;
int nmat ;
Material *mat ;
OPinfo* op;
OPSinfo* opsn;
int t_insnode;
int t_outsnode;
int global_t_nodes ;
int global_t_elms ;
int global_t_domains;
DomMesh *dom ;
int *off_gdom ;
} PartMesh ;
19. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
typedef struct {
double lambda ;
double c ;
double rho ;
double stfn ;
double inheat ;
} Material ;
typedef struct {
int num ; /* node number */
double val ; /* value */
} Bcond ;
typedef struct {
int num1 ; /* element number */
int num2 ; /* surface number */
double val ; /* outside temperature */
double alpha ; /* alpha */
} Bconv ;
typedef struct {
int num1 ; /* element number */
int num2 ; /* surface number */
double val ; /* outside temperature */
double emis ; /* emisssivity */
double fact ; /* geometric factor */
} Bradi ;
39. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Incomplete Balancing Domain Decomposition for Large Scale
Thermal-Solid Coupling Problems
Masao OGINO, Ryuji SHIOYA,
Hiroshi KANAYAMA, and A.M.M. MUKADDES
Kyushu University, JAPAN
WCCM VI in conjunction with APCOM’04, Beijing, China
Minisymposium: Large Scale Coupled Problems and Related Topics
- Proposition of the IBDD-DIAG method for large scale analysis
- Thermal-solid coupling analysis with ADVENTURE system
- Thermal-solid coupling analysis of 12 million nodes model
40. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
ADVENTURE System
Module-based architecture of ADVENTURE system.
Development of Computational Mechanics System
for Large Scale Analysis and Design
http://adventure.q.t.u-tokyo.ac.jp/
Open source CAE software for large scale analysis and design
Solver modules,
ADV_Solid
Elastic-plastic analysis solver
ADV_Forge
Rigid-plastic analysis solver
ADV_Impact
Impact-contact analysis solver
ADV_Thermal
Thermal conductive analysis solver
ADV_Fluid
Thermal flow analysis solver
ADV_Magnetic
Electromagnetic analysis solver
41. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Modules for thermal-solid coupling
• ADVENTURE_Thermal:
steady and non-steady thermal conductive analysis in solid,
HDDM(Hierarchical Domain Decomposition Method)*
solver,
BDD(Balancing Domain Decomposition) preconditioner.
• ADVENTURE_Solid:
static elastic, elastic-plastic and large deformation analysis,
HDDM solver,
BDD preconditioner.
*G.Yagawa and R.Shioya, Parallel finite elements on a massively parallel computer with
domain decomposition, Computing Systems in Engineering, 4:4-6, 1994, pp.495-503.
*
HDDM is one of the parallel FEM based iterative DDM.
42. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Flow chart of thermal-solid weak coupling analysis with
ADVENTURE system.
Mesh, Thermal properties,
B.C. for thermal analysis
Domain decomposition
by ADVENTURE_Metis
Thermal analysis
by ADVENTURE_Thermal
Thermal analysis
by ADVENTURE_Thermal
Structural analysis
by ADVENTURE_Solid
Structural analysis
by ADVENTURE_Solid
Mesh, Mechanical
properties,
B.C. for solid analysis
Domain decomposition
by ADVENTURE_Metis
Deformation,
stress, strain, ...
Temperature
informations
Thermal analysis flowThermal analysis flow Solid analysis flowSolid analysis flow
43. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Balancing Domain Decomposition (BDD)*
DDM,
which solve the interface problem using Preconditioned CG
(PCG) method.
Neumann-Neumann (N-N) preconditioner,
which works to a residual vector as a local subdomain
preconditioner.
Coarse grid correction,
which deletes a coarse space component from a vector.
*J.Mandel, Balancing Domain Decomposition, Communications
on Numerical Methods in Engineering, Vol.9, 1993, pp.233-241.
44. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Domain decomposition
Interface
Domain Decomposition Method (DDM)
gSuB =
fKu =
Original finite element problem:
Analysis domain
After eliminating interior DOF, we get a reduced
problem on interface, called the interface
problem,
local Schur
complement
( ) ( ) ( ) ( )( ) ( )i
IB
i
II
Ti
IB
i
BB
i KKKKS
1−
−=
Schur complement matrix( ) ( ) ( )∑= =
N
i
Tiii RSRS 1
( )
( )
( ) ( )iTi
i
i
:R
:f
:K
Ω→Ω restriction operator
local stiffness matrix
local RHS vector
I : interior dofs, B : interface dofs
where
Interior
45. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
N-N preconditioniner
( ) ( ) ( ) ( ) ( ) ( )
V
N
i
Tiiiiii IRDR,VV:D =∑→ =1
Partition of unity
:V The space of the dofs on
interface, and
The space of the dofs
on interface of Ωi.
( ) :V i
( )
( ) ( )
( ) ( )
=Γ
≠Γ
= −
−
φ
φ
i
U
i
i
U
i
i
:S
~
:S
S 1
1
† S(i)
is non-singular,
S(i)
is singular.
N-N preconditioning operator is
where
( ) ( ) ( ) ( ) ( ) ,RDSDRT N
i
iTiiii
NN ∑= =− 1
†
46. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Construction of the coarse space
Local coarse space is defined by,
( ) ( ) ( ) ,VZRangeSKer iii ⊂⊂
( ) :V i Space of the interface of subdomain i.
−
−
−
=
0100
0010
0001
12
13
23
xx
xx
xx
z )i(
X
For the elastic problem,
At point on ,)x,x,x(X 321
)i(Ω∂
.X:B
,ZBZ
)i()i(
X
X
)i(
X
)i(
X
)i(
Ω∂→
∑=
For the heat conductive problem,
[ ]T)i( ,...,Z 11= Num. of `1` is the num.
of interface dof for ( ) .iΩ
47. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
BDD preconditioner
Preconditioning matrix is,
( ) ( ) ( ) ( ) ( ) ( )[ ]NNN ZDR,....,ZDRR 111
0 =
000 SRRS T= : coarse grid operator
SRSRP T
0
1
00
−= : S-orthogonal projection operator
where
If you set solution of the coarse space problem to initial guess of
the interface problem, BDD preconditioner can be simplified,
( ) ( ) ,PITPIPSM
T
NNBDD −−+= −
−1
( ) .TPISM NNBDD −
− −=1Simplification
48. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
BDD-DIAG: BDD with diagonal-scaling
• The N-N preconditioner requires,
- generalized inverse matrices,
- large memory size to keep inverse matrices.
( ) ( )
( )
( )( ){ }i
V
ii SdiagmaxISS
~
i
⋅⋅+= α
a regularization:
low precision
Moore-Penrose pseudo-inverse: high computational costs
• The diagonal-scaling preconditioner is employed as
local subdomain preconditioner, named BDD-DIAG.
This method shows memory shortage and low computational costs,
while it increases the number of iterations compared with BDD.
( ) ( ){ } ( ) .RSdiagRT N
i
Tiii
DIAG ∑= =
−
1
1
( ) ,TPISM DIAGDIAGBDD −=−
−
1
where
49. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Incomplete balancing operator
• Coarse grid correction:
- to invert coarse grid operator
- the dimension of is equal to,
in thermal analysis:
in solid analysis:
It is difficult to apply for large scale problems.
• Incomplete parallel Cholesky factorized operator is employed for
coarse grid correction*
.
- may increase the number of iterations,
- may decrease the computational costs greatly,
total computation time is expected to be reduced.
Even if it corrects using an approximate solution, it is investigated
numerically that results are the same.
ntdisplacemerigidofdofN ×
N
0S
0S
50. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Incomplete balancing domain decomposition
000 RS
~
RS
~ T= : incomplete coarse grid operator
S
~
RS
~
RP
~ T
0
1
00
−= : S-orthogonal projection operator
where
IBDD,
IBDD-DIAG,
Because the incomplete coarse grid correction cannot delete a
coarse component, these methods cannot be simplified.
( ) ( ) ,P
~
ITP
~
IP
~
SM
T
NNIBDD −−+= −
−1
( ) ( ) ,P
~
ITP
~
IP
~
SM
T
DIAGDIAGIBDD −−+=−
−
1
51. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
BDD type preconditioners
IBDD: N-N and incompletely balancing,
IBDD-DIAG: Diagonal-scaling and incompletely balancing,
( ) ( ) .P
~
ITP
~
IP
~
SM
T
NNIBDD −−+= −
−1
( ) ( ) .P
~
ITP
~
IP
~
SM
T
DIAGDIAGIBDD −−+=−
−
1
( ) .TPISM NNBDD −
− −=1
( ) .TPISM DIAGDIAGBDD −=−
−
1
BDD: N-N and completely balancing,
BDD-DIAG: Diagonal-scaling and completely balancing,
52. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Numerical examples
Thermal-solid coupling problems:
• Pipe model with 0.1M nodes,
for comparison with BDD type preconditioners.
• ABWR model with 12M nodes,
as the large scale analysis on MPP.
A related example:
• Seismic response analysis of ABWR,
which is an application of IBDD-DIAG in solid analysis.
53. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Pipe model
FE Mesh (tetrahedron-quadratic)
Elements: 61,913
Nodes: 114,112
DOF in thermal analysis: 112,166
DOF in solid analysis: 339,411
Temperature distribution
Equivalent stress distribution
and deform configuration
The pipe model
100 C
P
X
X
X
X
X
20 C
54. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
0.0
5.0
10.0
15.0
20.0
25.0
30.0
35.0
40.0
40 120 200 280 360 480 640 800 1200 1600
Number of subdomains
Time[sec]
0
10
20
30
40
50
60
70
80
90
100
40 120 200 280 360 480 640 800 1200 1600
Number of subdomains
Numberofiterations
Comparison of BDD performances
in thermal analysis
# subdomain vs. # iterations # subdomain vs. computational time
BDD-DIAG
BDD-DIAG
IBDD-DIAG
IBDD-DIAG
IBDD IBDD
BDD
BDD
Pentium4 3.0GHz X 4
55. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Comparison of BDD performances
in solid analysis
0
50
100
150
200
250
300
350
40 100 160 220 280 340 400 1000 2000 3200
Number of subdomains
Numberofiterations
0
100
200
300
400
500
600
700
800
40 100 160 220 280 340 400 1000 2000 3200
Number of subdomains
Time[sec]
# subdomain vs. # iterations # subdomain vs. computational time
IBDD-DIAG
IBDD-DIAG
BDD-DIAG
BDD-DIAGIBDD
IBDDBDD
BDD
Pentium4 3.0GHz X 4
56. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
ABWR model
Advanced Boiling Water Reactor
FE Mesh (tetrahedron-quadratic)
Elements: 7,486,792
Nodes: 11,794,506
DOF in thermal analysis: 12 M
DOF in solid analysis: 35 M
FE mesh of the ABWRABWR (TEPCO, Inc.)
59. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Seismic response analysis of ABWR
using IBDD-DIAG
-300.0
-200.0
-100.0
0.0
100.0
200.0
300.0
400.0
0.0 4.0 8.0 12.0 16.0 20.0 24.0 28.0 32.0 36.0 40.0 44.0 48.0 52.0
Elapsed time [sec]
Acceleration
-300.0
-200.0
-100.0
0.0
100.0
200.0
300.0
400.0
0.0 4.0 8.0 12.0 16.0 20.0 24.0 28.0 32.0 36.0 40.0 44.0 48.0 52.0
Elapsed time [sec]
Acceleration
elapsed time: 0.0sec – 0.96sec,
Stress distribution and deformation of ABWR
1940 Elcentro, acceleration history
in NS direction, sec.t 020=∆
CPU: ES 1,024 processors
Time step: 200 (4.0sec)
Computational time: 6.97hr
CPU: ES 1,024 processors
Time step: 200 (4.0sec)
Computational time: 6.97hr
Time-integration: Newmark’s beta
Damping: none
60. 第 2 回 ADVENTURE 上級者セミナー
2005 年 9 月 16 日 ( 金 )
Conclusions
• A construction of new BDD type preconditioner, called
IBDD-DIAG.
• A thermal-solid coupling analysis with 12M nodes model on
MPP.
• Good performance of the IBDD-DIAG method in
computational time compared with original BDD for large
scale analysis.
In the future works,
• Comparison with other methods, AMG or FETI.
Editor's Notes
Domain Decomposition is non-overlapping type.
Then original system is reduced into the interface system.
This equation called the interface problem.
The coefficient matrix is called Schur complement.