View
215
Download
0
Tags:
Embed Size (px)
Citation preview
F.Yuasa at ACAT2002
Multidimensional Integration Package: DICE
andits parallelization
F.Yuasa / KEKK.Tobimatsu / Kogakuin Univ.
S.Kawabata / KEKACAT2002
24-28, Jun. 2002 at MSU, Moscow
F.Yuasa at ACAT2002
BASES
• Multidimensional Integration Package• Stratified and Importance sampling metho
d• Singular function can be integrated• Up to 100 dimensions• Heavily used in GRACE framework
F.Yuasa at ACAT2002
x
y
(x , y)
Y
X
(X , Y)
When singularities go along a diagonal line,we need appropriate variable transformation.
F.Yuasa at ACAT2002
DICE• Developed by K.Tobimatsu and S.Kawabata
– First version of DICE in 1992– Research Reports of Kogakuin Univ. No.72
(1992)• Divide the integral region into 2Ndim hypercubes• Two kinds of sampling method• DICE Input
– Ndim, Expected Error, # of Sampling points, Maximum division level, Maximum # of iteration
F.Yuasa at ACAT2002
Level = 2 Level = 3
How to divide HypercubeNdim=2
Regular sampling and random sampling
regular
F.Yuasa at ACAT2002
Example 1
I1 dx1dx2
2x2
(x1 x2 1)2 20
1
0
1
F.Yuasa at ACAT2002
F.Yuasa at ACAT2002
Example 2
I2 dx1dx2
x22(1 x1
2 x22)
(x12 x2
2 a2)2 2 1
1
1
1
F.Yuasa at ACAT2002
F.Yuasa at ACAT2002
Example 3
I3 dx1dx2dx3
x12 x2
2 x32(1 x1
2 x22 x3
2)
(x12 x2
2 x32 a2)2 2
1
1
1
1
1
1
F.Yuasa at ACAT2002
F.Yuasa at ACAT2002
Example 4
I3 dx1dx2dx3dx4 f (1 x1
2 x22 x3
2 x42)
(x12 x2
2 x32 x4
2 a2)2 2 1
1
1
1
1
1
1
1
,
Rx12 x2
2 x32 x4
2,
and
f (R) (R 1/4)(R 1/2)2(R 3/4)
4
F.Yuasa at ACAT2002
f (R) (R 1/4)(R 1/2)2(R 3/4)
10 1
a0.8
10 2
a0.8
F.Yuasa at ACAT2002
Results of I4Package Eps =10**(-1) Eps =10**(-2)
DICE-mpi (3.1408
+-0.0029)E-02
(-1.0662
+-0.0011)E-02
ParInt1.1 0.031216955
+-0.0016138
-0.010710149
+-0.00055466
BASES (3.174411
+-0.047391)E-02
(-1.056945
+-0.035239)E-02
Analytical
results
0.0314313 -0.0106773
F.Yuasa at ACAT2002
Results of I4 (part2)Package Eps =10**(-3) Eps =10**(-4)
DICE-mpi (-1.6070
+-0.0023)E-02
(-1.6676
+-0.0024)E-02
ParInt1.1 -0.01593609
0.01293747
-0.0114787661
0.0206822789
BASES (-1.614249
+-0.009806)E-02
(-1.639091
+-0.033053)E-02
Analytical
results
-0.0160761 -0.0166246
F.Yuasa at ACAT2002
• More complicated Integrand
• # of dimensions = 4
• # of lines in FORTRAN = about 300 lines
Example 5
F.Yuasa at ACAT2002
Results of Example5Package Result # of Sample
points
DICE-mpi1 processor
(1.0638+-0.0011)E-13 24798768
ParInt1.1
1 processor
(1.0622+-0.0529)E-13 2100000110
BASES (1.064086+-0.000337)E-13
9996350
F.Yuasa at ACAT2002
Results of Example5 (part2) Package Result # of Sample
points
DICE-mpi1 processor
( -1.1529+-0.0019)
E-13
24798768
ParInt1.1
1 processor
(-1.1452+-0.0652)E-13 2100000110
BASES (-1.154076+-0.000681)E-13
9996350
F.Yuasa at ACAT2002
Results of Example5 (part3) Package Result # of Sample
points
DICE-mpi1 processor
( -8.8675+-0.0872)E-15 0.98%
33956096
ParInt1.1
1 processor
We did not try
BASES We did not try
F.Yuasa at ACAT2002
Parallelization
• We use MPI for the parallelization.
• Parallelization is useful for higher dimensional integrand
• Parallelization is useful for complicated integrand
• Example 5 is calculated by the parallelized DICE
F.Yuasa at ACAT2002
Speed Up
# of CPUs 1 2 4 8
CPU time
[sec]
3232.80 1648.40 839.00 431.68
Speed Up 1.00 1.96 3.85 7.49
Example5
F.Yuasa at ACAT2002
Summary
• We have developed DICE.
• DICE is available to Vector Processor.
• DICE is available to Parallel Processor.
• We have used MPI for parallelization.
• For the complicated integrand, parallelization shows good scalability.