Escolar Documentos
Profissional Documentos
Cultura Documentos
H H$G
q
H
x
G
!x
H
, i"1, . . . , n (2)
While ordinarily the solution of this problem requires O(n`) operations, for large n, one can
compute the forces with the FMM in O(n) operations only. We outline the FMM in the next
paragraph. For details, the reader is referred to References 36.
The FMM exploits the well-known idea that the potential eld induced by a cluster of charges
can be approximated by the potential eld of their total charge or higher-order moments, as long
as the target charges are suciently far away from the source charges. In the FMM, the clusters
are formed via hierarchical partitioning of the unit cube into 8", 8, 8`, etc. identical cubic cells
which interact among themselves based on the machinery of the multipole and solid spherical
harmonic (local) expansions. The computation proceeds in three stages. First, the multipole
coecients (charge, dipole, etc.) are computed for all cells. Second, the cellcell interactions are
computed so that the local expansion coecients of the smallest cells are related to the multipole
coecients. Third, the potentials are computed by adding the potentials due to the local
expansions and nearby charges. As a result the solution is expressed as
u(x
G
)"
'N>'
`
I
l
I
h
I
(x
G
!c)#
H$G
q
H
x
G
!x
H
#error (3)
and
f
G
"
'N>'
`
I
l
I
[!q
G
Vh
I
(x
G
!c) ]#remaining terms (4)
FAST SOLUTION METHOD FOR THREE-DIMENSIONAL MANY-PARTICLE PROBLEMS 1217
1998 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 42, 12151229 (1998)
In these equations, ls are the local expansion coecients, hs are the solid spherical harmonics,
c is the centre of the smallest cell containing q
G
, and the tilde sign denotes that the second sum
includes nearby charges only. The upper limit of the local expansion is ( p#1)` because this is the
number of the solid spherical harmonics of order less or equal to p. The error term, which can be
bounded a priori, is controlled by choosing a suciently large p. Note that most of the FMM
machinery is needed for computing l sevaluation of (3) and (4) for given l s is a simple
computation performed pointwise, one charge at a time.
In order to extend the FMM to many-body problems associated with discretized integral
equations of three-dimensional linear elasticity we introduce two generalizations of the many-
body electrostatics problem. The rst generalization is related to the fact that collocation BEMs
are endowed with two point sets. One point set is formed by the nodal points which are also used
as the collocation points. The other point set is formed by the quadrature points. To distinguish
between these two sets we regard the collocation points as the target points and the quadrature
points as the source points. In what follows, we denote the target points by x and the source
points by y. The second generalization is related to the fact that the fundamental solutions of
linear elasticity involve the biharmonic kernel x!y to which, without modications, the FMM
cannot be applied. In this regard, we observe that only minor modications are required in order
to extend the FMM to problems in which (i) the charges at the source points are prescribed in
terms of a function q"Q( y) and (ii) the forces at the target points are prescribed in terms of
a linear map F on u(x) so that
f
G
"F[u(x
G
), x
G
]"
' N>'
`
I
l
I
F[h
I
(x
G
!c), x
G
]#remaining terms (5)
It is obvious that (5) is a pointwise computation that can be implemented once the FMM is
supplemented with a pre-processing procedure for evaluation of q"Q(y) and a post-processing
procedure for evaluation of F[h
I
(x), x]. Note that with the introduction of the generalized
many-body electrostatics problem we amplify an observation of Greengard that the FMM is
a separation of variables technique for the harmonic kernel x!y. From this perspective, we
extend the FMM to the kernel X(x)( y) x!y.
As an example, let us consider how the generalization we have introduced can be applied to the
biharmonic kernel. In the following expression, we use the subscripts to denote Cartesian
components of x and y:
x!y" x) x
GHI
F
(x)
;
1
x!y
; 1
GHI
Q
( y)
#(!2x
)
GHI
F
`
(x)
;
1
x!y
; y
GHI
Q
`
(y)
#(!2x
`
)
GHI
F
`
(x)
;
1
x!y
; y
`
GHI
Q
`
(y)
#(!2x
`
)
GHI
F
"
(x)
;
1
x!y
; y
`
GHI
Q
"
(y)
# 1
GHI
F
`
(x)
;
1
x!y
; y ) y
GHI
Q
`
( y)
In this representation the harmonic kernel is distilled from the biharmonic one and as a result the
many-body problem associated with the three-dimensional biharmonic equation is reduced to
ve generalized many-body electrostatics problems.
1218 Y. FU E A.
Int. J. Numer. Meth. Engng. 42, 12151229 (1998) 1998 John Wiley & Sons, Ltd.
In conclusion of this section, we observe that, strictly speaking, the FMM is not an O(N)
method because in order to optimize the performance the partitioning depth should be a logarith-
mic function of N. We avoid this issue in this paper and keep referring to the FMM as an O(N)
method; for details see Reference 7.
3. GOVERNING EQUATIONS
The principal objective in this section is to formulate the boundary integral and corresponding
algebraic equations of many-particle problems in a form suitable for the FMM application. We
accomplish this objective by combining well-established BEM procedures` with new representa-
tions for the fundamental solutions of three-dimensional linear elasticity.
3.1. Integral equations
In this section, we formulate the boundary integral equations for a simple many-particle
problem that involves an innite binder with M identical particles perfectly bonded to the binder.
This problem is selected for demonstration purposes. There are no conceptual obstacles to
extending our approach to problems that involve nite binders, dierent particles, and debon-
ding.
The problem of interest is formulated in three steps. First, we formulate an auxiliary boundary-
value problem for a nite body with M embedded particles. Second, we formulate the boundary
integral equations for the nite body and introduce new representations for the fundamental
solutions. Finally, we derive the boundary integral equations for the innite body. From now on,
we use subscripts to denote Cartesian components and superscripts to denote attributes. For
position vectors, we use interchangeably the indicial and direct conventions; for example, the
notations x and x
G
are equivalent.
Consider a nite three-dimensional linear elastic body formed by a binder D and M embedded
identical particles DK perfectly bonded to the binder (Figure 1). We denote the surface of this body
Figure 1. A binder containing identical particles
FAST SOLUTION METHOD FOR THREE-DIMENSIONAL MANY-PARTICLE PROBLEMS 1219
1998 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 42, 12151229 (1998)
by I", the interface boundaries by IK, and their union, l+
K
IK, by I. Lames constants of the
binder are denoted by z and j and the corresponding Poissons ratio by v. For the particles, the
corresponding constants are marked with the asterisk superscript. The body is loaded by
a displacement eld u"
G
(x) prescribed on I".
The displacement eld in the body, uJ
G
(x), is the solution of the following boundary-value
problem:
( z#j)uJ
H GH
(x)#juJ
G HH
(x)"0 x3D (6)
( z*#j*)uJ
H GH
(x)#j*uJ
G HH
(x)"0 x3l+
K
DK (7)
uJ
G
(x)"u"
G
(x) x3I" (8)
|uJ
G
(x)""0 and |tI
G
(x)""0, x3I. (9)
In (9), | . . . " denotes the jump along the normal vector, n
G
(x), as dened in Figure 1.
Equations (6)(9) are equivalent to a system of M#1 integral equations. The rst equation in
this system is dened on the binder boundary, I"I, and each of the remaining equations is
dened on the boundary of a particle:
1
2
uJ
G
(x)!
"
GH
(x, y)uJ
H
(y) dy!
GH
(x, y) uJ
H
(y) dy"!
"
GH
(x, y) t
I
H
(y) dy!
GH
(x, y) t
I
H
(y)
(10)
and
1
2
uJ
G
(x)#
*
GH
(x, y)uJ
H
(y) dy"
*
GH
(x, y)t
I
H
(y) dy, m"1, 2, . . . , M (11)
The change in sign between (10) and (11) is due to the fact that the normal vector on the matrix
boundary points to the particle. In (10) and (11) and in the rest of the paper it is implied that x is
an arbitrary point on the surface on which the integral equation is dened.
The kernels
GH
(x, y) and
GH
(x, y) introduced in (10) are the fundamental solutions of three-
dimensional linear elasticity. Usually they are expressed as
GH
(x, y)"
1
16j(1!v) x!y
(3!4v) o
GH
#
(x
G
!y
G
)(x
H
!y
H
)
x!y`
(12)
and
GH
(x, y)"
1!2v
8(1!v) x!y`
[!(x
G
!y
G
)n
H
(y)#(x
H
!y
H
) n
G
(y)#o
GH
(x
Q
!y
Q
)n
Q
(y)]
#
3
8(1!v) x!y`
(x
G
!y
G
) (x
H
!y
H
) (x
Q
!y
Q
) n
Q
(y) (13)
where o
GH
is the second rank identity tensor. The kernels *
GH
(x, y) and *
GH
(x, y) introduced in (11)
can be obtained from (12) and (13) once j and v are replaced by j* and v*, respectively.
1220 Y. FU E A.
Int. J. Numer. Meth. Engng. 42, 12151229 (1998) 1998 John Wiley & Sons, Ltd.
The kernels
GH
(x, y) and
GH
(x, y) can be also expressed as
GH
(x, y)"P
GH
(x)
1
x!y
#Q
G
(x)
1
x!y
y
H
(14)
and
GH
(x, y)"R
GHN
(x)
1
x!y
n
N
( y)
#S
GN
(x)
1
x!y
n
N
( y)y
H
(15)
where
P
GH
(x)"
1
16j(1!v)
[(3!4v)o
GH
!x
H
c
G
] (16)
Q
G
(x)"
1
16j(1!v)
c
G
(17)
R
GHN
(x)"
1
8(1!v)
[(1!2v) (o
HN
c
G
!o
GH
c
N
)!2(1!v)o
GN
c
H
#x
H
c
G
c
N
] (18)
and
S
GN
(x)"!
1
8(1!v)
c
G
c
N
(19)
In (16)(19), the symbol c
G
denotes the partial derivative with respect to x
G
. Note that in (14) and
(15) the terms dependent on x are separated from those dependent on y except for the harmonic
kernel x!y.
Now let us evaluate (10) and (11) under the conditions that I" tends to innity and u"
G
(x)"
c"
GH
x
H
, where c"
GH
is a constant tensor. To do this we rst consider an auxiliary problem for the nite
binder from which the particles are removed and in which the linear displacement eld
u"
G
(x)"c"
GH
x
H
is realized by prescribing u"
G
(x)"c"
GH
x
H
on the entire boundary I"I. The corres-
ponding stress and traction elds are denoted by o"
GH
and t"
G
, respectively; t"
G
"o"
GH
n
H
. The integral
equation for the auxiliary problem parallels (10):
1
2
u"
G
(x)!
"
GH
(x, y)u"
H
(y) dy!
GH
(x, y)u"
H
( y) dy"!
"
GH
(x, y) t"
H
(y) dy!
GH
(x, y) t"
H
(y)
(20)
Once (20) is subtracted from (10), the integrals over I" disappear as I" tends to innity and we
obtain
1
2
u
G
(x)!
GH
(x, y) u
H
(y) dy#
GH
(x, y)t
H
(y) dy"0 (21)
where
u
G
(x)"uJ
G
(x)!u"
G
(x) and t
G
(x)"t
I
G
(x)!t"
G
(x) (22)
FAST SOLUTION METHOD FOR THREE-DIMENSIONAL MANY-PARTICLE PROBLEMS 1221
1998 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 42, 12151229 (1998)
In terms of the perturbation elds, (11) is rewritten in the form
1
2
u
G
(x)#
K
*
GH
(x, y)u
H
(y) dy!
K
*
GH
(x, y) t
H
(y) dy"g
G
(x) (23)
with
g
G
(x)"!
1
2
c"
GH
x
H
!c"
HQ
K
*
GH
(x, y) y
Q
dy#o"
HQ
K
*
GH
(x, y)n
Q
( y) dy (24)
In (23) and (24) it is implied that m runs from 1 to M.
In the remainder of the paper, we are concerned with solving integral equations (21) and (23).
The most challenging part of this task is to solve (21) which can be characterized as a coupled
problem for M particles. For this problem the use of the FMM is critical. In contrast, (23)
represents M relatively small problems coupled only through (21). For these problems, a conven-
tional BEM is adequate.
3.2. Algebraic equations
In this section, (21) and (23) are converted into algebraic equations following standard nite
element discretization techniques.
The nite element representation of I is based on the map
xI
G
()"
J
xIJ
G
J () (25)
In (25) is the position vector in the master element space, xIJ is the position vector of the node
l of the element k, and J () is the shape function of the node l. For isoparametric nite elements,
the maps uI
G
[x()] and tI
G
[x()] mimic (25),
uI
G
[x()]"
J
uIJ
G
J () and tI
G
[x()]"
J
tIJ
G
J () (26)
When (25) and (26) are substituted in (21) we obtain
1
2
u
G
(x)!
I J
GH
[x, y()]uIJ
H
J ()J[y()] d#
I J
GH
[x, y()]tIJ
H
J ()J[y()] d"0
(27)
In (27) the integrals are over the master element space G and J[y()] is the Jacobian correspond-
ing to (25). With collocation at the nodal points xO, followed by numerical integration of (27), we
obtain
1
2
u
G
(xO)!
I J Q
GH
[xO, y(Q)]uIJ
H
J (Q) J[y(Q) ]cQ#
I J Q
GH
[xO, y(Q) ]tIJ
H
J (Q) J[y(Q)]cQ"0
(28)
1222 Y. FU E A.
Int. J. Numer. Meth. Engng. 42, 12151229 (1998) 1998 John Wiley & Sons, Ltd.
In (28) s runs over the quadrature points of the master element and cQ is the integration weight at
Q. Upon substitution of (14)(19), (28) can be rewritten as
1
2
u
G
(xO)!R
GHN
(xO)
I JQ
1
xO!y(Q)
uIJ
H
J (Q) n
N
[y(Q) ]J[y(Q) ]cQ
!S
GN
(xO)
I J Q
1
xO!y(Q)
uIJ
H
J (Q) n
N
[y(Q) ]y
H
(Q) J[y(Q)]cQ
#P
GH
(xO)
I J Q
1
xO!y(Q)
tIJ
H
J (Q) J[y(Q)]cQ
#Q
G
(xO)
I J Q
1
xO!y(Q)
tIJ
H
J (Q) y
H
(Q) J[y(Q)]cQ "0. (29)
Again, note that in (29) the terms dependent on x are separated from those dependent on y except
for the harmonic kernel x!y.
The discretized forms of (23) and (24) are similar to (28):
1
2
u
G
(xO)#
Y
I J Q
*
GH
[xO, y(Q) ]uIJ
H
J (Q) J[y(Q) ]cQ
!
Y
I J Q
*
GH
[xO, y(Q)]tIJ
H
J (Q) J[y(Q)]cQ"g
G
(xO)
and (30)
g
G
(xO)"!
1
2
c"
GH
xO
H
!c"
HN
Y
I J Q
*
GH
[xO, y(Q) ]yIJ
N
J (Q) J[y(Q)]cQ
#o"
HN
Y
I J Q
*
GH
[xO, y(Q) ]nIJ
N
(Q)J[y(Q) ]cQ (31)
The wide hat implies that the sums involve the nodes and quadrature points that belong to
a single particle.
At this stage, it is useful to rewrite (29)(31) in a symbolic matrix form. To this end we form
N-dimensional nodal vectors u, t, and g organized on a particle-by-particle basis, so that each
vector is formed by M blocks. Consequently, the operators acting on u and t are represented as
N;N matrices formed by M;M blocks. Thus, we rewrite (29) as
Au#Bt"0 (32)
and (30) and (31) as
A*u#B*t"g (33)
Equations (32) and (33) can be combined into
(A!BB*A*) u"!BB*g (34)
or simply
Cu"b (35)
FAST SOLUTION METHOD FOR THREE-DIMENSIONAL MANY-PARTICLE PROBLEMS 1223
1998 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 42, 12151229 (1998)
Also it is useful to rewrite (35) in the residual form
r(u)"Cu!b"0 (36)
The structure of the matrices in (32) and (33) reects the physics of the problem. The matrices
Aand Bare dense and their o-diagonal blocks represent pairwise particle interactions while the
diagonal blocks represent self-interactions. The o-diagonal blocks become negligible for well-
separated particles, and their decay rates follow directly from (12) and (13). The matrices A* and
B* are block-diagonal because the particles interact with each other only through the binder and
these interactions are included in the matrices A and B. For identical particles, the diagonal
blocks in each matrix are the same.
4. ITERATIVE SOLUTION STRATEGY
In this section, we outline a procedure for solving (35) using O(N) memory and O(N) operations.
This procedure consists of three components: (i) decomposition of u along the residual Krylovs
vectors r, Cr, C`r, C`r, etc., (ii) matrix-vector multiplication with the FMM, and (iii) precondi-
tioning.
The solution vector u is obtained following the generalized minimum residual algorithm
(GMRES) of Saad and Shultz.` This algorithm requires construction of the residual Krylovs
vectors. The kth Krylov vector v'I' is constructed by induction:
v'I'"Cv'I' with v'"'"r
Following (34), this multiplication is performed as
v'I'"Av'I'!Bw'I' (37)
with
w'I'"B*A*v'I'
Observe that the term w'I' can be computed in O(N) since both A* and B* are block-diagonal.
Now comes the crucial step application of the FMM to (37). To explain this step let us go back
to (29) and replace there the nodal displacements with the nodal values of v'I' and the nodal
tractions with the nodal values of w'I'. The four sums in (29) have the same structure and
therefore it is sucient to explain how to apply the FMM to the rst one. To connect this sum
with the generalized many-body electrostatics problem dened in Section 2, we identify the
collocation points xO with the target points, the quadrature points y(Q) with the source points, the
expression in the curly brackets with the generalized charge Q[y(Q)], the sum with the potential
u(xO), and the operator R(xO) with the generalized force operator F[u(xO), xO]. Observe that the
generalized charge is a second rank tensor and therefore it has to be represented by nine ordinary
charges. Thus, the rst sum is computed by invoking the FMM nine times. For the remaining
sums, the FMM is invoked three times for the second and third sums, and once for the last sum,
so that the total number of times the FMM is invoked is 16.
For large algebraic problems, like those corresponding to many-particle problems, precon-
ditioning is important because it allows one to reduce the number of required iterations.
Preconditioning involves multiplication of both sides of (35) by a matrix P, the preconditioner,
which is as close to C as possible. We choose P to be block-diagonal, with each block being
equal to the inverse of the corresponding diagonal block of C. Formally, this preconditioner is
1224 Y. FU E A.
Int. J. Numer. Meth. Engng. 42, 12151229 (1998) 1998 John Wiley & Sons, Ltd.
known as block-Jacobi. For many-particle problems, the block-Jacobi preconditioner has a clear
physical meaningit exactly represents non-interacting particles. Therefore the matrix PC
represents the inter-particle interactions only. Details of incorporating the preconditioner into
the GMRES can be found in Saad and Shultz.`
5. PARALLEL IMPLEMENTATION
The goals for our parallel implementation of the FMM algorithm were eciency and scalability
and provision for mesh adaptivity in a second phase of algorithm and code renement. These
goals were attained by building the parallel code on the Scalable Dynamic Distributed Array
(SDDA) data management infrastructure."` The SDDA provides normal array semantics for
any array of objects (where each may have a dierent structure) which is distributed across
multiple processors and which may expand and contract with preservation of locality and
without copying to follow adaptive mesh renement. The SDDA was originally developed to
support hp-adaptive nite element methods"` for which it provides data management for the
geometry data of the adapting nite element model. This requires support for ecient handling of
dynamic arrays of small objects which is the fundamental data management problem for the
parallel FMM implementation.
The implementation of the SDDA is based upon use of space-lling curves and separation of
address management (indexing) from storage management.' Space-lling curves` map the
three-dimensional physical space to a one-dimensional space. Several authors have previously
used space-lling curves for implementation of parallel tree codes.`