You are on page 1of 12

SUBSTRUCTURING TECHNIQUES-STATUS

AND PROJECTIONS?

AIMED K. NOOR,*
HUSSEIN A. KMEL! and ROBERT E. FULTON~

(Received March 19n; received for publication 1 June 1977)

A&net-Status and recent developments of substructuring techniquesand their applicationto structuralanalysis


and design are summarized. Discussion focuses on a number of aspects including: multilevel substructuring
algorithms; use of hyperrnatrix and other sparse matrix schemes, use of substructuring in automated design
systems and application of substructuring in elasto-plastic probkms. Numerical exampks are presented to
demonstrate the reduction in the number of arithmetic operations and disk storage requirements obtained by using
multilevel substructuring techniques. Also discussed are the potential benefits of using such techniques with new
computing hardware such as CDC STAR-100 and minicomputer systems.

NowwcLATlmR external forces at constrained and external


[C] connection matrix defined in eqn (8) nodes delined by qn (12)
{C) vector of undetermined coefficients de&d in incremental external forces defined by eqn (27)
qn (22) reduced incremental load vector defined by
(CPU),,(CPU)d CPU times required for substructuring and qn (30)
direct analysis, respectively (Fig. 13) structure load matrix defined by qn (7)
d,“, d: original and modified set of design variables reduced substructure and structure load vec-
[I] identity matrix tors defined by qns (24) and (26)
&I. i&J. displacemeot unknowns at internal and inter-
substructure. stifncss matrices defined by qn face nodes defined by qn (I)
IK,l. [&,I I (1) displacements at constrained and external
[&I reduced (effective) substructure stiffness nodes, qn (It)
matrix defined by qns (3) and (4) incremental displacements defined by qn (27)
I&l*I&It matrix of displacements at all interfaces, qn
(7)
stiffness matrices of co&9ra&l substructure
Wcfl.
[&I‘ disk storage requirements in substructuring
defined by qn (12) and direct analysis, respectively (see Fig. 6)
L!Ll.[!L]?
[K&l,[K-l I disk storage requirements for direct analysis
[&I* I&J with one node per block (see Fig. 7)
1 incremental stiffness matrices detined by qn number of design variables
tmnsfommtion matrices defined by eqns (13)
reduced incremental stiffness matrix defined in and (17)
qn (2% matrices of reduced base vectors for internal
structure stiffness matrix defined by qn (7) and interface unknowns defined by qns (20)
reduced substructure and structure stiflness and (21)
matrices deflned by eqns (23) and (24)
beginning address of submatrix [K,] (see Pii IrvmomlcTIoN
21 The concept of dividing a structure into parts to carry
number of floating-point multiplications in
substructuring and direct analysis, respec-
out analysis dates back to the precomputer era and is
tively (see Fig. 6) closely related to matrix partitioning. With the advent of
number of floating-point multiplications in digital computers, substructuring technology has been
direct analysis with one node per block (see further developed and applied to several stress analysis,
I%. 7) eigenvalue and automated structural design problems.
number of substructures For example, several finite element computer programs
external forces at internal and interface nodes, now have substructuring as a standard capability (e.g.
defined by qn (1) NA!3TRAN, ASKA, SESAM-69, DAISY and GIFTS,
reduced (effective) substructure load matrix
Refs.[l-51). Substructuring applications to eigenvalue
defined by qn (5)
problems are surveyed in Refs. [6-91.
Substructuring developments, however, have been
oriented primarily to capabilities provided by current
tPresented at the Second National Symposium on Com- third generation computers. Recent and expected
puterized Structural Analysis and Design of Engineering and developments in computers offer potential for improved
Applied Science, George Washington University, Washington, capabilities in calculations for the future. Such new
D.C. 29-31 March 1916.
systems include not only large advanced computers such
tProfessor of Engine&n8 and Applied Science, George
Washington University Center at NASA-Langley Research Cen_
as the STAR-100, ILLIAC IV and ASC but also
ter, Hampton, VA 23665, U.S.A. numerous minicomputers which provide dramatic reduc-
OProfessor of Aerospace Engineering, University of Arizona, tion in computing costs coupled with a high degree of
Tucson, Arizona. man/machine interaction. Both classes of computers are
1Monager. IPAD Project Ot?lcc. NASA-Langley Research well suited for subsaucturing and it is appropriate to
Center. Hampton, VA 23665, U.S.A. review substructuring technology to identify methods
621
622 A. K. Noon et al.

particularly well suited for this evolving computer The internal degrees of freedom [U,‘i’] are expressed
environment. This paper summarizes the status and in terms of interface degrees of freedom [U,“‘] and are
recent developments in static substructuring techniques eliminated from eqn (1) giving (superscript j is omitted
and their application to structural analysis and design. for convenience):
Since the subject is very broad, discussion focuses
herein on a number of aspects which are of interest to [&I = - [Kil-‘[&I [ubl + [Kiil-‘[pil, (2)
the authors and which have not been fully covered in the
literature. These aspects include: multilevel substructur- and
ing algorithms, use of hypermatrix and other sparse
matrix schemes, substNcturing techniques for automated [&I [Ubl = rs1 (3)
design systems and substructuring techniques on new
computing systems such as the CDC STAR-100 and where [&I is the effective substructure stiffness matrix,
minicomputer systems.
[&I = iKbbl- [K&l[&-‘[&I9 (4)
SUMMARY OF SUBSt’RUCTUlUNGMETHODSIN
SrATIC SFBUCTURAL. ANALe3Is
and [&,I is the effective SubstNcture load matrix,
The method of substructuring for static structural
analysis is based on subdividing the large structure into
IsI = lpbl - LKbil[Kiil-‘[PiI. (5)
smaller parts which are analyzed separately to obtain
relationships between forces and displacements at the The matrix inversion in eqns (4) and (5) is not normally
part interfaces. These interface variables are then done explicitly but the products [Kii]-‘[Kib] = [Ml] and
determined and the results used to obtain the unknowns [K,]-‘[Pi] = [Q] are obtained by solving:
within each substructure. Whilst substructuring tech-
niques can be applied with the force or mixed methods
IKiilIM j Qil = [Kb \ pil (6)
of structural analysis (see, e.g. Refs.[lO, 1l]), the dis-
cussion herein is limited to the displacement method. An via decomposition, forward reduction and back sub-
historical account of the development of substructuring stitution.
techniques is given in Ref. [12] where it is shown that Matrix [I&] can also be obtained by reducing the
these techniques are closely related to matrix stiffness matrix of eqn (1) to an upper triangular form
partitioning[ 131 and that Kron’s method of tearing[l4] using Gaussian elimination but terminating elimination
represents the first computational procedure of the when the final row of the rectangular matrix [Kii Kib]
modern substructuring analysis. This section gives a has been reduced. The matrix [&I then occupies the
brief review of the mathematical theory of substructur- location of the original matrix [&,,I. This method is
ing for static structural analysis together with some discussed in Ref. [lo] and is referred to as partial tri-
recent developments relative to sparse matrices, hyper- angulation.
matrices and introduction of constraints. The governing equations for the whole structure are
obtained by assembling the stiffness and load matrices
Review of static substructuring theory [&I and [Pb] of the m substructures. The equations can
For a structure composed of m substructures (Fig. I), be written in the following form:
the well-known force equilibrium equations for the jth
substructure (j = 1 to m) are: P-l [%I = [PI (7)

in which [%I is the matrix of displacements at all inter-


faces; and [YLJand [g] are the stiffness and load matrices
of the whole structure. The displacement matrix [U,,] for
where [Ui”qV [Q,(“], [P,oq and [P,o’] are displacement each SubStNctttre is extracted from [%I and substituted
and external forces at the internal and interface nodes, into eqn (2) to yield internal displacements and forces.
respectively; the [K]‘s are stiffness submatrices and In cases where the sequential order of the boundary
superscript j refers to the jth substructure. displacement components is not the same in all sub-
structures, or when different coordinate systems are used
for different substructures the assembly process can be
simplified by introducing a connection matrix, which
relates [II&“‘]to [%I, namely:

[V,“‘] = [C”‘] [%I. (f-v

The corresponding expressions for the structure stiffness


and load matrices are:

WI = g, c-TuGO’l b?‘l (9)

ISI = 2 [C”‘l=[&.o)]. (10)


i-1

Fig. 1. Origiial structure and a substructure. A hierarchy of substNctures (multilevel substructur-


Substructuring techniques-status and projections 623

ing) can be obtained by combining (assembling) goups (11) are applied in a recursive manner, the equivalence
of substructures and eliminating the internal unknowns between multilevel substructuring and nested dissection
within them. This process amounts to applying eqns can be established. The equivalence between nested dis-
(l)-(lO) in a recursive manner. When the lowest (first) section and multilevel substructuring suggest the follow-
level substructure consists of only few elements, arran- ing consequences:
ged in a standard pattern, it is usually denoted a super 1. The rules for numbering the nodes to reduce the
element, i.e. a substructure may be thought of as consist- number of arithmetic operations and the fill-in using
ing of a number of super elements. The te.rm multilevel nested dissection scheme[lS, 161can be used in dividing
super element is sometimes used 10 refer to multilevel the structure into substructures and in numbering the
substructuring[31. nodes of individual substructures. This latter aspect is
particularly important since poor numbering of the nodes
Relationship between substructuring and sparse matrix of individual substructures can offset the computational
techniques advantages of substructuring techniques.
Recent work has shown that the banded, profile and 2. If all substructures are different, then the number of
wave front schemes are not the most efficient techniques arithmetic operations involved in the nested dissection
for solving a system of algebraic equations[lS, 161where scheme is identical with those involved in the substruc-
efficiency is measured in terms of the number of arith- turing technique, provided of course, that the internal
metic operations and fill-in (number of zero terms in the nodes are numbered in the same way. On the other hand,
original equations which become nonzero in the solution if a number of substructures are identical, then the
process). References115 and 161 conclude that more substructuring technique involves significantly less
efficient solution techniques are based on an ordering arithmetic operations than the nested dissection scheme.
strategy of the fundamental unknowns called nested dis- This situation is true even when the substructures are
section which results in a highly sparse nonbanded differently oriented with respect to the global (structure)
matrix of the equations. axes in which case the effective substructure and load
The basic idea of nested dissection is to divide the matrices have to be transformed to the global axes.
nodes of the complete structure into sets of internal and These transformation usually involve considerably less
interface nodes. Each set of internal nodes is further arithmetic than that involved in the generation of
subdivided into internal and interface nodes of the effective substructure matrices.
second level. The process is continued until it is no
longer possible to divide the internal node sets. The Hypermatrir storage schemes
numbering of the nodes starts with the node sets of the The hypermatrix (or block matrix) scheme based on
highest level. The internal nodes of any level are num- partitioning the structural matrices in both the row and
bered before the interface nodes of that level. The zero column directions[l8] has proved to be a versatile and
elements in the resulting algebraic equations are neither effective technique for handling large systems of equa-
stored nor operated upon in the process of solution. tions resulting in the finite element analysis of large and
A comparison between sparse matrix schemes and complex structures. The location of the nonzero sub
substructuring techniques is given in [VI. In this section matrices in the hypermatrix is identified by an address or
the equivalence of multilevel substructuring and nested a pointer matrix (see Fig. 2). A zero entry in the addresc
dissection schemes is discussed, and in succeeding sec- matrix denotes a zero submatrix which is neither stored
tions numerical examples are presented to demonstrate nor operated upon. Among the advantages of the
the reduction in arithmetic operations and disk storage hypermahix scheme are the independence of the central
requirements obtained by using multilevel substructuring memory requirements from the problem type and size, a
techniques. controllable ratio between CPU and I/O times and a high
Assume the structure is broken into m substructures. modularity.
If all the equilibrium equations corresponding to the m The use of hypermatrix storage schemes in conjunc-
substructures are grouped together, one obtains: tion with substructuring techniques has the additional
I-

[K;;‘] ,/”
p?

VT’1 = pcm’
h ;

[C(“]T(K6:‘] . . . [C’Z’]T[K:‘j.. . (C’m’]=[K;y’] 3 rm’rP,o”J


(11)

where eqn (8) has already been used to express the advantage of allowing exploitation of the sparsity of the
interface unknowns for the individual substructures substructure matrices in the analysis, an important fea-
[&,“I in terms of the structure unknowns (matrix of ture for complicated substructures. Moreover, the use of
displacements at all interfaces) [%I. The same set of hypermatrix storage scheme appears to hold promise for
eqns (11) results from the nested dissection approach[lS] use on fourth-generation computers (e.g. CDC STAR-
and numbering the internal nodes for all of the m sub 100) where a large central memory is available for the
structures first followed by the interface nodes. If eqns analysis. The identification of the substructure shapes
624 A. K. Noon et al.

KI.15.2 0 0 Q 5.”

+Yl
1cr
.2 K2.3 0 Kzn-1 K2.n
o INIERNALNODES
5.3 K3.4 0 0 0 CONSTRAINEDNODE
OEXlERNALraE
K4.4 0 K4.n

~SYMtAETRIC
-5.
K
. 1 ‘“qj
il-l.ll-l
O
K
"." ADDRESS IPOINTERI
MAMTRIX

HYPERMATRIX Ii,/
= BEGINNINGADDRESS
OF SUMATRIX KLj

Fig. 2. Hypermatrix and address (pointer) matrix. Fig. 3. Representative constrained substructure model.

and levels can be accomplished by proper numbering of (13) has been used):
the nodes. The governing equations for individual sub-
structures are written (in the format of eqns 11) or (14)
backing storage before starting the reduction process.
The use of substructuring techniques on the STAR-100 where
computer is discussed in succeeding sections.

Constrained substructuring technique


The constrained substructuring technique such as that
proposed in Ref.[19] is based on applying kinematic
constraints to the boundaries of individual structures to
reduce the number of independent interface degrees of
freedom. It is closely related to the hybrid finite element
models developed by Pian[20], wherein a displacement
interpolation scheme is assumed along the boundaries.
The constrained substructuring procedure provides the
experienced analyst with additional flexibility in model-
ing, by which reasonably accurate solutions can be
obtained with considerable savings in computer time. It
is feasible to incorporate into this procedure a library of
kinematic constraints such as rigid links, and linear,
quadratic or spline interpolations. The constrained substructuring technique not only has
A schematic representation of a constrained substruc- computational advantages but also simplifies assembly of
ture model is shown in Fig. 3. The interface nodes are elements with different dimensionality (e.g. shell ele-
divided into external nodes, designated by open squares, ments assembled to one-dimensional beam elements, and
and constrained nodes, designated by solid circles. The beams tied to membrane elements, see Fig. 4).
constrained nodes are made to be kinematically depen-
dent on the external nodes. The internal nodes are COMPUTER RESOURCEUTILUATION
designated by open circles. The number of external Computer costs and data management
nodes is usually selected to be less than the number of The size and topo!ogy of substructures selected
the constrained nodes. significantly impacts the amount of data being managed
The equilibrium equations of the jth substructure, eqn and required computer resources such as: central pro-
(1), are now rewritten in the following modified form cessing unit (CPU), rapid access (central) memory and
(where, for convenience, superscript j has been deleted). peripheral processing unit (PPU). The CPU cost is pro-
portional to the number of arithmetic operations and
central memory cost is a function of the core space
requested and time occupied. The PPU cost includes the
costs of disk storage and input/output (I/O) channel
utilization. Such costs have different weights depending
where [UC], [UJ, [PC] and [P,] are substructure dis- on system type and degree of use. For example, a system
placement and external forces at constrained and exter- with an extremely fast disk drive may not penalize I/O
nal nodes, respectively. operations as much as one with a slow drive. A system in
The kinematic relationship between the constrained which most users do not require extensive computation,
and external interface displacements can be written as: encourages the use of the central processor-so that it
becomes an inexpensive resource.
[I-J,1= [reel [Kl (13) For static structural analysis, substructuring com-
putations can be divided into three phases[211: (a)
where [r,] is the transformation matrix. generation of basic substructures; (b) reduction and
The effective (reduced) substructure equations are assembly of substructure matrices and (c) local analysis
obtained by eliminating [Vi] and [UC] from eqn (12). The of individual substructures. The computer resources
resulting equations have the following form (where eqn most used in each step of these three phases are listed in
625

- CfflTRM.

kiif
-----DATA FLOW

SEAMELEMENT
0 MTERNAL NODES
* CONSTRAINEDNUDES

Fig. 4. Examples of constrainedsubstructuremodels. Fig. 5. Sequence of operations in an example two-level sub-


structuringanalysis.

Table 1, and Ref. [22] describes a computer program to model assembly and solution to proceed in a certain
identify the number and size of substructures that results order (see Fig. 5 for the sequence of operations in an
in the least cost in a given computing system. example two-level substructuring problem).
The data management aspects in a general purpose
finite element program with substructuring capability are Simulator generated numerical examples
more important than in direct analysis. Data associated To assess the computational advantages of substruc-
with the problem is usually stored in a data base com- turing techniques, a computer program was developed to
posed of many distinct and often complex files. For simulate the substructuring analysis process. No actual
example, a substructure is a finite element model of a floating-point computations were carried out but
structure and requires the complete set of data files measurements were made of computing resource use. In
necessary for model description. The substructure, particular, the following information was determined:
however, also requires storage of the files associated 1. Number of floating-point arithmetic operations
with its formation (such as the condensed stiffness, load (multiplications, additions and divisions),
and mass matrix files and intermediate files used during 2. Number of I/O transfers to and from central
the process of static condensation) and alphanumeric memory and the number of words transferred and
identifiers to allow the program, during the main analysis, 3. Disk storage use which is measured by the space
to access the substructure files for automatic assembly of required for the substructure and structure stiffness
substructure matrices. Furthermore, the constrained matrices (equal to the space required by the decomposed
substructure approach also requires storage of additional matrix on the left-hand side of eqns (11)).
quantities such as kinematic constraints and connec- A large number of problems were solved using both
tivities. In multilevel substructuring, cross references direct analysis (zero-level substructuring) and various
must also be provided between a primary level structure levels of substructures. In all cases the hypermatrix
and subsequent lower level substructures to permit storage scheme was used to exploit the sparsity pattern

Table 1. Computerresourcesused in each step of the static substructuringanalysis

Analysis step Computer resources


most used
Phase I - Generation of Basic Substructures

Computation of elemental stiffness matrices CPU


Assembly of elemental stiffness matrices I/O
Load application I/O and possibly CPU
Formation of boundary constraint matric’es I/O
(for constrained substructures)

Phase II - Reduction and Assembly of


Substructures
Reduction of substructure matrices and Both CPU and I/O
transformation to global coordinates
Assembly of substructure matrices I/O
PhrrseIII - hCti tiJ’SiS Of SubStNChWS6

Computation of elemental displacements I/O


and stresses

CA.3Vd. 8, No. S-F


626 A. K. NOORet al.

Table 2. Number of arithmeticoperationsand disk storagerequirementsin the direct and substructuringanalyses

Direct T Substructuring analysis


Grid* analysis One level Two levels Three levels Four levels
I
a) Number of floating point multiplications
8X8 1,016,739 715,203 722,763 __-_____ ________
16 x 16 11,160,94? 7,362,803 6,543,731 6,462,515 -_______
32 x 32 143,987,859 90,378,387 68,953,347 58,190,715 57,490,Oll
b) Number of floating point additions
8X8 1,002,321 703,377 711,009 ________ ________
16 x 16 11,067,617 7,289,057 6,476,897 6,396,833 --__--__
32 x 32 143,321,985 89,864,577 68,529,249 57,821,265 57,127,761
c) Disk storage requirements (in words)
6X6 28,836 23,652
16 x 16 186,660 147,492
32 x 32 1,331,748 1,027,820
and 32*Number
x 32 grids,
of degrees of freedom = 486, 1734 and 6534 for the 8 x 8, 16 x 16,
respectively.

of the substructure and structure matrices. Each of the A$ = NUMBER OF WLTIPLICATIONS SD = DISK STORAGE REQUIREMENTS
IN DIRECT ANALYSIS FOR DIRECT ANALYSIS
matrices was partitioned in the row and column direc- WITH ONE NODE PER WITH ONE NODE PER
tions (see Fig. 2). Zero blocks were neither formed nor BLOCK BLOCK

operated upon. However, to simplify the analysis the


zeros within each block were not exploited, since
exploitation of these zeros would require additional
bookkeeping and complicate comparisons.
The effect of variation of the block size on the number
of arithmetic operations and storage requirements was
investigated and typical results are presented in Table 2
and in Figs. 6-8 for square grids of 8 x 8, 16x 16 and
32 x 32 elements having six degrees of freedom at each
node. While the substructures are geometrically similar, 0 2 4 6 8 0 2 4 6 S
they are assumed to have different stiffness matrices. NUMBER OF NODES PER BLOCK NUMBER OF NODES PER BLOCK
The results presented in Fig. 6 and Table 2 show that,
Fig. 7. Effect of variationsin the size of submatrices on arith-
except for the small 8 x 8 grid, as the number of sub- metic operations and storage requirements.
structuring levels increases, the number of arithmetic
operations and disk storage requirements decrease.
However, the cost of data handling and automatic mesh
generation is likely to increase. These latter drawbacks
are more than offset by the other well-known advantages
of substructuring techniques such as breaking the prob-
lem into smaller manageable pieces, simplifying input
and output and aiding management of large projects (see,
for example, Refs. [22,231). Figure 7 shows the effect of
variation of the block size from 6 x 6 (corresponding to

Ss= DISK STORAGE


REQUIREMENTS
IN SUBSTRUCTURING
ANALYSIS ANALYSIS

MULTlPLlCATlON
RATIO

Fig. 8. Schematic representation of the structure equations.

the degrees of freedom at one node) to 48 ~48 (cor-


responding to the degrees of freedom at eight nodes). As
I I h I I can be seen from Fig. 7, if the sparsity pattern of
NUMBER Of SUSSlRUClURING tEk5 NUMSiR OF S:SSTRUiNR,,iG lEV& individual blocks is not exploited, the increase in the
Fig. 6. Effect of numberof levels of substructuringon floating- block size results . _. _ in increasing both the arithmetic
point arithmetic operations. .and_ disk storage requirements. operations and disk storage requirements. However, the
Squaregnds of elements. I/O transfers are reduced. Figure 8 shows a schematic of
Substructuring techniquecstatus and projections 627

the equations before and after decomposition for an 8 x 8 sitivity analysis vectors, measuring the rates of change of
grid. The two cases of direct analysis and two-level the response variables with respect to the design
substructuring are shown. In the latter case the equilib- variables; dp are the original design variables (I = I to s)
rium equations for the individual substructures are and an asterisk refers to a modified quantity. The sen-
grouped together in a manner similar to that of eqns (I 1). sitivity analysis vectors can be obtained by first differen-
The reduction in the fi/f in obtained by using substruc- tiating the substructure equations, eqns 1, with respect to
turing techniques is noted in the table in Fig. 8. These design variables, eliminating the internal unknowns, and
computational gains obtained by substructuring are assembling the different substructures in exactly the
similar to those reported in Refs. 123,241for more com- same manner as in the original substructure analysis (see
plex structures. Ref.[30]). The computational cost of obtaining the sen-
The computational advantages of substructuring tech- sitivity analysis vectors is only a fraction of the cost of
niques are strongly dependent on the selection of the size the full analysis.
of individual substructures and the numbering of their A large number of numerical experiments on statically
nodes. A number of attempts have been made to indeterminate trusses and wing-box finite element models
rationalize the selection of substructures (see, e.g. have shown that the Taylor series expansions provide
Refs.[25-271 which only treat simple structures). For highly accurate approximations for (UY} and (IQ} for
guidelines regarding the numbering of the nodes in sim- moderate changes in the design variables, provided the
ple structures, reference can be made to nested dis- design variables are chosen as reciprocals of the sizing
section schemes[lS, 161. A heuristic approach for the variables. For large modifications in the design variables
selection of substructures for complex structures in the reduced basis technique described subsequently was
order to minimize the cost of the analysis is presented in found to be more appropriate.
Ref. [27]. Reduced basis technique. In this technique each of the
displacement vectors (I-J:} and (U:} are approximated
by a linear combination of s linearly independent sen-
Numerous applications have been made of nonlinear sitivity analysis vectors normalized to reduce numerical
mathematical programming techniques to the optimum roundoff error, in which s is the number of design
design of small and medium-sized structures. For large variabks and is much smaller than the dimensions of the
structural systems, however, excessive computer substructure stiffness matrices. This process is expressed
requirements limit the usefulness of such techniques. A by the following equations:
number of techniques have been proposed in recent
years to reduce computational effort. Among the more
promising approaches which use substructuring concepts (19)
are: (a) use of multilevel optimization techniques, and (b)
use of approximate reanalysis techniques with substruc- where
turing.
The first approach is based on decomposing the struc-
ture into substructures each with its own objective func- ~~~l=[[~},[~]....[~}] (20)
tion and constraints (see, e.g.[28,29]), the vector of
design variables is partitioned into substructure and in-
teraction variables. The optimization problems for the
~~*I=[[~},[~]....[~]] (21)
individual substructures are tirst solved independently
and then the interaction solutions are obtained itera- and (c},., is a vector of undetermined participation
tively. Different optimization techniques can be used for coefficients and a bar denotes a normalized vector.
individual substructures and the interaction problem. The governing equations for (c) may be obtained by
Success with this method has been limited and further applying the stationary potential energy principle-after
work is needed to facilitate its usefulness. expressing the total potential energy in terms of the trial
The second approach which can be used in conjunc- displacements of eqns (19).
tion with, or independent of, the tirst approach is to use
reanalysis techniques to generate the solutions for the
mod&d structure (corresponding to a modified set of
design variables) in significantly less computer time than in which
that required to solve the full structure equations. Two
efficient reanalysis procedures which have potential in
substructuring are the Taylor series expansion and
WY = cm ICI (2%
reduced basis techniques, a discussion of each follows.
Taylor series expansion. In this technique the new
displacement vectors {Uf} and {Ut} for each substruc-
ture corresponding to a mod&d set of design variables
d? (I= 1 to S) are approximated by a truncated Taylor’s
series

(Pr}= [+ilil’{Ptl
+ [@blr(pbl (26)
where{Wddd and (W,/~dJ are the first-order sen- and m is the number of substructures.
628 A. K. Noon ef al.

The major savings obtained by using the reduced basis elastic substructures are reduced only once prior to the
technique are: (a) avoiding the decomposition of the start of the nonlinear analysis. On the other hand, the
modified matrices [K$]; (b) reducing the assembly of the plastic substructures are mod&d during load incremen-
[K*,] and {Pr} matrices to straight forward matrix addi- tation and equilibrium iterations. Such an approach can
tion; and (c) significantly reducing the size of the struc- be mathematically described by the following set of
ture matrix [YL’$,..The reduced basis technique appears incremental equilibrium equations (see Refs. [32,33]).
to be well suited for problems where the number of
design variables has been reduced by design variable
linking, a process by which some design variables are
expressed in terms of others (see Ref. [311).The accuracy
of the two reanalysis techniques for a transmission tower where the K’s are incremental stiffness matrices, AU’s
(Fig. 9) is given in Fig. 10. For example, a reduction of and AP’s are incremental displacements and external
50% in the cross sectional areas of the vertical members forces, and subscripts e and p refer to the elastic and
of substructure IV, results in less than 35% change in the plastic substructures, respectively.
maximum displacement of the tower. The Taylor series If AU, are eliminated from eqns (27) one obtains:
expansion and reduced basis techniques calculate this
displacement to within 0.7% and 0.2%, respectively.

suBsrRtJcTtJRtNGIN ELAsTcMLAsrK where


PROBLEMS
In elasto-plastic problems where yielding is confined to ]$,I - W,,l [~eeI-‘weJ
= W,,l (2%
a small portion of the structure (e.g. crack propagation
problems), computations can be substantially reduced by {Ah,,= W’pl- [&,I KJ’W’J. (30)
reducing the size of the nonlinear problem. This can be
accomplished by dividing the structure into two sets of [&,] is the reduced incremental stiffness matrix and
substructures. The first set consists of those substruc- {AP,} is the reduced incremental load vector. Note that
tures which are known to remain elastic throughout the only part of the stiffness matrix, namely [K,,.] and [Z&l,
analysis. The second set consists of the substructures needs to be modified at each load increment. The sub-
which may undergo yielding during the analysis. The matrix [K,] which is much larger than [Z&J does not
have to be reformulated or decomposed at succeeding
load increments. Such a procedure was used in
Refs. 134and 351 to study crack extension and closure
under cyclic loading.

hnNtcoMPuTgR-MDEDstlWrRuclwRAL
ANALYSIS
The recent advances in minicomputers provide a new
dimension to cost effective computer-aided structural
analysis and design. Such minicomputers may be pur-
chased for less than $100,000, operate in a normal office
environment, and carry out selected structural analysis
and design tasks for only a fraction of the cost for the
same task done on a large-scale computer (see, e.g.
Ref.[36]). However, minicomputers do not, in general,
replace large-scale computers, but provide a less power-
ful, but inexpensive, adjunct to such capability.
Fig. 9. Substructuringanalysisof a transmissiontower (219bars). Reference[36] shows substantial benefit in the use of
minicomputers for such tasks as data management, in-
teractive graphics support and structural analysis of
moderate-sized problems. Furthermore, some organiza-
tions having purchased a minicomputer will likely seek to
stretch its capabilities to very demanding activities
and/or classes of problems.
This section discusses some potential minicomputer
applications for substructuring both as an independent
computer and as an adjunct to a large computer. To
identify the cost-effective features of the mini, Table 3
+I\ I IN SUBSTRUCTURE IV
I gives comparative costs of conducting basic arithmetic
tasks on three computers: a minicomputer without float-

lot
‘jy:I
D 50 -5D -10 0 IO
+ MODlFlCATlON OF A5
30 x)
ing point hardware (PDP 15 at the University of
Arizona-Tucson), a minicomputer with floating point
hardware (PDP 15), and a large-scale computer (CDC
6400 at the University of Arizona). While the large
computer results are dependent on specific installation
and user demands at the time, the results, nevertheless,
Fig. 10. Accuracy of maximum displacement obtained by the indicate that for many activities the minicomputer, both
reanalysistechniques. with and without floating point hardware, can be a cost-
Substructuringtechniques-status and projections

Table 3. Comparisonof unit cost of arithmeticoperationson largeand minicomputers

Item CDC 6400


(Univ. of Arizona)

Floating point Addition 1.1 x 10’6


arithmetic
time (sec.) Multiplication 5.7 x 10’6
Number of Addition 0.45 0.06 I 1.11 ~ I
operations
per second Multiplication 0.09 0.06 0.74
sr;F

effective tool for many computational tasks. Although process of (1) substructure decomposition, (2) main sys-
the mini is slower, the fact that it may operate in a tem analysis and (3) detailed analysis of the specific
dedicated mode means that the turn-aroundtime may be substructure are shown in Table 4. These results indicate
shorter than with a heavily loaded large computer. The that it is practical to conduct substructure analysis for
minicomputer may be used for moderately large struc- such moderate-sized structures using a minicomputer
tural analysis problems through use of substructuring. and the total time of about 1 hr and 11 min is not unac-
The following discusses a specific example. ceptably large. Furthermore, a floating-point processor is
Consider the uniformly thick plane stress problem available on the market which would significantly reduce
shown in Fig. 11 which represents a metal plate per- this time.
forated at equal intervals and subjected to in-plane load- The computer times in Table 4 indicate that almost 50
ing. Two static load cases are considered: (I) inertial and 19 min. respectively, were required for steps (1) and
loading in the horizontal direction; and (2) inertial load- (2), while only about 3min were required for step (3).
ing in the vertical direction. The structure is separated The results also show that the major time consumers
into 16 substructures as shown with each substructure (denoted with an asterisk) are reduction activities for the
having 200 elements with two in-plane degrees of substructure and decomposition for the main system
freedom at each node. Because of the repetitiveness, analysis. These tasks are essentially independent and can
only one constrained substructure was formed and its be separated in a modularized structural analysis such as
stiffness reduced to eight main analysis node points done here. This situation suggests that future structural
around its boundary. Normal displacements on a boun- analysis programs should be organized so that large,
dary between two external substructure nodes are con- calculation-oriented tasks are carried out on a large-scale
strained to a cubic displacement function whereas tan- computer well suited for number crunching, and assembly
gential displacements are assumed to vary linearly. processes, which are primarily data management and
Loads were applied to the substructure nodes and could be thought of as pre- or post-processing, are car-
automatically condensed and assembled for the system ried out on a minicomputer. Furthermore, the calculation
analysis. The overall system structural analysis was car- phases, such as those noted with the asterisk, are pri-
ried out followed by a detailed analysis of a represen- marily vector-oriented and may be well suited for a
tative substructure at the bottom right corner. Figure 12 future distributed computing environment where the
shows overall deflections for the second load case, as large computer is a vector-processing computer such as
well as deflection and stress results for the representative the STAR-1001371.
substructure. Such distributed computing concepts which would
Minicomputer times required for the three step facilitate independent calculations on different machines

NUMBER M cmFlJTATlm TIME 011


SUBSTFZUCTURES- 16 PDPQ IwlnKRJ1F.P. MRDWARQ
NUMBER OF D.O.F. IN EACH
SUBSTRUCNRE = 41) l SUBSTRLlCNRE 49 min 46 set
NUMBER m D.OF. IN MIN REDUCTIDN
ANAL% IS = 255 l K4 IN SWIM lB*nZOsa
AMLVS IS
l LOCAL SUBSTRUCTURE 3 tin 4 I*:
MJALYSIS
TOTAL 1 hr I1 min 4 WC
yt

ELEWENTIAVOUT x WD REPRESENTATION x RIBHTSUBSTRUCNRE DEFLECTEDSUBSTRLICTURE

Fig. It. Substructuremodel of a perforatedplate. Fig. 12. Stress and deflectionresults for perforatedplate model.
630 A. K. NOOR et al.

Table 4. Minicomputer times for analysis of perforatedplate


Structural Description
No. of active nodes in each constrained subatmcture = 240
No. of degrees of free&m per node = 2
Type of boundary interpolation = 8 cubic segments/substructure
No. of constrained substructures = 16
No. of nodes in main analysis = 85
No. of degrees of freedom/node = 3
No. of loading cases = 2
Computation Times on the PDP-15 Computer (without Floating PoM Hardware)
Substructure Decomposition (Total for all substructures)
Bandwidth optimization 1 min. 57 sec.
l Stiffness assembly 12 min. 48 sec.
+ Decomposition of internal nodes 8 min. 05 8ec.
* Stiffness and load condensation 26 min. 50 sec.
Total 4%min. 40 sec.
Main System Analysis
Bandwidth optimiznkion 49 sec.
Stiffness asssembly 3 min. 33 sec.
+ Decompoeition time 8 min. 50 sec.
Back substitution 5 min. 8 sec.
Total 18 min. 20 sec.
Detailed Subatructuw Analysis (Total for all substructures)
ComputaUon of aubetructure deflection8 1 min. 30 sec.
Computation of sub&ucture stresses lmirLS4lWc.
T&al SIX&L 4sec.

*Major time-consuming operatione.

are still evolving and software needs to be developed to To assess the effect of using multilevel substructuring
support routine data transfers among minicomputers, on the CPU time and storage requirements on the STAR,
large computers, and vector processors. Nevertheless, a computer program was developed to simulate the sub
these results suggest that future finite element work structuring analysis process and estimate the CPU time
should be oriented toward development of finite element and storage requirements (see Table 5 for the scalar and
substructure analysis methods in highly modular form vector timings on the STAR). The three square grids
with standardized procedures for module input/output so (8 x 8, I6 x I6 and 32 x 32) considered in the preceding
that specific tasks, such as those noted in Table 4, can be sections were analyzed. In each case the hypermatrix
assigned to the most appropriate computer hardware. It storage scheme was used to exploit the sparsity pattern
also suggests that the concept of pre- and post-process- of the substructure and structure matrices. The opera-
ing should be redefined to encompass a vast array of tions on submatrices were vectorized in the manner
specialized data management oriented tasks at the described in Ref.[38]. Zero blocks were neither formed
beginning or end of each modular analysis and that such nor operated upon. However, the zeros within each
tasks be assigned where possible to a minicomputer as block were not exploited. To exploit these zeros the
the most cost-effective. timely way of conducting these sparse vector capability of the STAR can be used, but
tasks. since this capability is still evolving, no timing infor-
stJBsmucruRlh.cON CLK SrAR-1m mation is available for sparse vector operations, and
COMPUTER therefore, to simplify the comparison, sparse vectors were
The fourth generation CDC STAR-100 computer has not considered.
many new features which distinguish it froin current Typical results are shown in Figs. 13(a)and (b). &ure
third-generation hardware. Features which appear to 13(a) shows the ratio of the estimated CPU time in the
influence most significantly finite element analysis in substructuring analysis to that in the direct analysis. For
general and substructuring techniques in particular are small problems, the reduction in the CPU time obtained
the vector and sparse vector pipeline processing capabil- by substructuring is small due to the use of short vectors.
ity and the virtual memory. These concepts have been As the problem size increases, the reduction in CPU time
discussed in Refs.[37,38] and are summarized in becomes more pronounced-up to a point, after which
Appendix A. Some numerical results based on the very small gain in CPU time is dbtained. Figure 13(b)
STAR-100 computer configuration at the NASA-Langley shows the ratio of storage requirements in substructuring
Research Center are given in this section. analysis to that in the direct analysis. As is clear from
Substructuringtechniques-status and projections 631

ICPl+ = CPU TIME means of exploiting some of the features of new computer
REQUIRE0 FOR
SURSTRUCNRING
hardware. For example
ANALYSIS SUBSTRUCTURING (a) Minicomputers are eminently suited for carrying
out the prelpost processing and data management tasks
in the substructuring analysis.
(b) The virtual memory of the CDC STAR-100 com-
puter is particularly effective for use with substructuring.
However, for large problems, data management costs are
expected to lead to less levels of substructuring on a
STAR-100 than on third-generation computers.
4. Reanalysis procedures used in conjunction with
substructuring can substantially reduce computational
0 P 0 4 requirements for automated (optimum) design of large
NU:BER 0: dS Gf N”hhER 0: LEVEL: OF
SUBSTRUCTURING SUBSTRUCNRING
structures.

Fig. 13. Effect of number of levels of substructuringon CPU REFZRENCEf


time and storagerequirementson the STAR.
I. The NASTRAN TheoreticalManualLevel 16.0, NASA SP-
221(03)(1976).
2. E. Schremand J. R. Roy, An Automaticsystem for kinema-
this figure, substructuring reduces storage requirements. tic analysis, ASKA, Part 1. Proc. of the IUTAM Symposium
The numerical results presented in this subsection sug- on High Speed Computing of Eiastic Structuns, Vol. 2,
gest that, for computational efficiency on the STAR, Universityof Liege, pp. 477-507 (1971).
substructuring should be used for large problems and 0. Egelandand P. 0. Araldsen,SESAM-69-A general pur-
should not be. carried out as far as on the third generation pose finiteelement program.Comput.Struct.4.41-68 (1974).
computers. This is because the data management cost is H. A. Kamel and A. Elbatouti, DAISY Engineer’s Manual.
Volumes 1and 2, Universityof Arizona,Tucson (1970).
likely to increase with the increase in the number of H. A. Kamel and M. W. McCabe, A graphics oriented
levels of substructures. interactive finite element time sharing package, GIFTS.
ResearchReport,Universityof Arizona,Tucson (1973).
6. R. Bamford, B. K. Wada, J. A. Garba and J. Chisholm,
CONCURWNG ltEMMtW Dynamicanalysis of largestructuralsystems. In Synthesis of
Status and recent developments of substructuring Vibrating Systems. ASME Booklet, pp. 57-71 (1971).
techniques and their application to structural analysis 7. W. A. Bentield, C. S. Bodley and G. Murrow,Modal syn-
and design are summarized. A number of substructuring thesis methods. Space Shuttle Dynamics and Aemelasticity
Work Group Symposium on Substructuring, NASA Marshall
aspects are considered herein including: multilevel sub-
Space FlightCenter, Alabama(1972).
structuring algorithms, use of hypermatrix and other
8. R. M. Hintz, Analytical methods in component modal syn-
sparse matrix schemes, use of substructuring in thesis. AlAA J. 13(No. S), 1007-1016(1975).
automated design systems and substructuring in elasto- 9. R. R. Craig, Jr. and C. J. Chang, A review of substructure
plastic problems. coupling methods for dynamic analysis. In Advances in
The results of the present study suggest several con- Engineering Science, Proceedingsof the 13 Annual Meeting
clusions relative to assessing and improving the of Engineenitg Science, Hampton,Va. NASA CP-2001,Vol.
efficiency of substructuring techniques and exploiting 2, pp. 393-408. 1-3 Nov. 1976.
new computer hardware advancements. 10. J. S. Przemieniecki and P. H. Denke, Joining of complex
1. Assessing the efficiency of different substructuring substructuresby the matrix force method. J. Aircmft 3(3),
236-243 (1966).
techniques for large structures can be effectively and
11. A. K. Noor and S. J. Hartley, Nonlinear shell analysis via
economically carried out through simulation of the mixed isoparametric elements, Comp. Struct. 7, 615-626
software operations, a technique well used in computer (1977).
science but not widely applied to engineering software. 12. R. H. h&Neal and C. W. McCormick,Computerizedsub-
Such an approach provides the key information on rela- structure analysis. Proc. of the World Congress on Finite
tive merits of alternate techniques without undue com- Element Methods in Structuml Mechanics, Dorset, England,
mitment of computer resources. (12-17 Oct. 1975).
2. The computational efficiency of substructuring 13. S. U. Benscoten, The partitioningof matrices in structural
analysis can be enhanced through a number of guidelines analysis. I. Appl. Mech. 15,303307 (1948).
14. G. Kron, Solving highly complex elastic structuresin easy
such as
stages. J. Appl. Mech. 22,235-244 (1955).
(a) Proper selection of the size, topology and node 15. J. A. George, Nested dissection of a regularfinite element
numbering of individual substructures. Improper node mesh. SIAM I. Numer. Anaivsis lOf2). 345-363 0973).
numbering can offset any computational advantages of 16. J. A. George, Numerical -experiments using dissection
substructuring. For guidelines regarding the numbering methods to solve n by n grid problems. Research Report
of nodes to reduce the arithmetic operations and disk CS-75-07,Departmentof ComputerScience, University of
storage requirements reference can be made to nested Waterloo,Waterloo,Ontario, Canada(1975).
dissection schemes. 17. F. W. Williams,Comparisonbetween sparse stiffnessmatrix
(b) Use of constrained substructuring techniques and substructuremethods. Int. J. for Numer. Methods in
Engng 5.383394 (1973). Discussion by K. Bemhardtand J.
wherein constrained nodes are kinematically coupled to
R. Roy, a(4), 601-602(1974).
external nodes. The use of those techniques can enhance
18. G. von Fuchs, J. R. Roy and E. Schrem,Hypermatrixsolution
modeling flexibility, reduce computer time and simplify of large sets of symmetricpositive-definitelinear equations.
assembly of elements with different dimensionality (for Computer Methods in Appl. Mech. Engng 1, 197-216(1972).
example shell elements connected to beam elements). 19. H. A. Kamel, D. Liu, M. W. McCabeand V. Phihppopotdos,
3. Substructuring techniques provide an effective Some developments in the analysis of complex ship struc-
632 A. K. NOORet al.

tures. In Advances in Computational Methods in Structural


Mechanics and Design, pp. 703-726. (Edited by J. T. Oden, SUMMNtY OF SFAR-100 CAl’ABlLWlES
R. W. Clougb and Y. Yamamoto), The University of Virtual memory
Alabama in Huntsville Press (1972). The memory organization on the STAR-100computer includes
20. T. H. H. Pian and P. Tona. Rationalization in deriving a magnetic core storage and random access auxiliary storage
element stiffness matrix by assumed stress approach. Pro; (CDC 819 disk storaae unit). Memorv is divided into 512 word
of the 2nd Conf. on Matrix Methods in Structural Mechanics pages where each word is.64 bits (approximately 15 decimal
AFFDLTR-68-140, pp. 44149 (1%8). digits). Also allowed are large pages each consisting of 128small
21. E. I. Field, D. N. Herting, D. L. Herendeen and R. L. pages or 65,536 words. Central memory 524,288 words (1024
Hoesly, The automated multi-stage substructuring system for small pages or eight large pages) with a capability to expand to
NASTRAN. In NASTRAN: Users’ Experiences, NASA TM twice that size. The CDC 819 unit is a moving disk system with a
X-3278,pp. 571-591(1975). capability of 33 million words on each of five drives (a total of
22. E. Klutz and S. Utku, Best partitions of a structure in a given 165million words). The central memory and disk are available to
computing environment. Presented at the Second National the user through virtual memory concepts. However, to perform
Symposium on Computerized Analysis and Design, George the computations, data must be stored in (or transferred to)
Washington University, Washington, DC. (29-31 Mar. 1976). central memory. The transfer of data to central memory is done
23. P. 0. Araldsen, The application of the superelement method in blocks of small or large pages which is considerably slower
in analysis and design of ship structures and machinery than computation time; hence, several computations per data
components. Presented at the National Symposium on item are needed to balance the CPU and I/O times. Moreover, to
Computerized Structural Analysis and Design, George avoid excessive I/O transfers during any particular phase of
Washington University, Washington, D.C. (27-29 Mar. 1972). computation, data references must be localized in one area of the
24. R. E. Miller, Jr., et al., Recent advances in computerized user’s memory space. Substructuring provides an effective way of
aerospace structural analysis and design. Presented at the achieving this data locality.
Second National Symposium on Computerized Structural
Analysis and design, George Washington University, Pipeline processing
Washington, D.C. (29-31 Mar. 1976). The central processor of the CDC STAR-188 computer con-
25. J. Lestingi and S. Prachuktam, A blocking technique for tains a pipeline arithmetic unit which segments arithmetic com-
large scale analysis. Comp. Struct. 3.669-714 (1973). putation into a sequence of basic operations (such as exponent
26. S. Utku, Systematic substructuring. J. Struct. Div. ASCE comparison, coefficient alignment, add and normalize shift). The
lOl(ST), 717-738 (1975). arithmetic unit can perform basic operations simultaneously on
27. M. O’Rourke and T. Bizal, Optimal substructuring schemes. independent pairs of data elements, each pair at a dilferent stage
Proc. of the National ASCE Structural Enaineeting Con- in the computation. A complete arithmetic operation requires a
ferenceon Methods of Structural Analysis, pp. 1037-1048. pair of data elements to stream through the entire pipeline (e.g.
Madison, Wisconsin (22-25 Aug. 1976). eight segments).
28. U. Kirsch. Multilevel aonroach to o&mum structural design. There are two types of data elements on STAR: scalars (a
I. Struct. Div. ASCE ii, 957-974 (1975). single-valued data element usually contained in one storage
29. U. Kirsch, M. Reiss and U. Shamir, Optimum deisgn by word), and vectors (an array of scalars usually contained in
partitioning into substructures. J. of the Struct. Div. ASCE consecutive storage words and addressable by a starting address
98.249-267 (1972). and length). If the same operation is performed on successive
30. A. K. Noor and H. E. Lowder, Approximate reanalysis data pairs (STAR vectors), results can occur at 40 psec intervals
techniques with substructuring. J. Struct. Div. ASCE for some operations[38]. These vector operations can lead to
101(ST8)1687-1698(1975). substantial gain in speed over sequential scalar arithmetic. Each
31. L. A. Schmit, Jr. and H. Miura, Approximation concepts for STAR vector operation has a fixed delay or startup time before
efficient structural synthesis. NASA CR-2552(1976). the first result emerges from the pipeline, and this effect on the
32. B. Aamodt, P. G. Bergan and H. F. Klem, Calculation of computation time decreases with increasing vector length. Exe-
stress intensity factors and fatigue crack propagation of cution time for a vector instruction is of the form:
semi-elliptical part-though surface cracks. Proc. of the
Second internaGona1 Conference on Pressure Vessel Tech- T=S+Np
nology, San Antonio, Texas, pp. 911-921 (1973).
33. B. Aamodt and P. G. Bergan, Numerical techniques in linear where T = total time, S = startup time, p = time per result and
and non-linear fracture mechanics. Computational Fracture N = vector length.
Mechanics, ASME 199-216(1975). Designing a program to make effective use of STAR pipeline
34. H. Armen, Jr., Application of a substructuring technique to capabilities requires substantial vectorization of tasks. For a task
the problem of crack extension and closure. NASA CR- to be vectorizable, it should have: (a) repeated operations; (b)
132458(1974). independence of each result from the others and (c) members of
35. J. C. Newman, Jr. and H. Armen, Jr.. Elastic-plastic analysis each operand packed into contiguous memory locations.
of a propagating crack under cyclic loading. AIAA 1. 13(g)
1017-1023(1975). Sparse vector capability
36. 0. 0. Storaasli, On the role of minicomputers in structural The STAR computer also has hardware features which are
design, Comp. Struct. 7, 117-123(1977). convenient for sparse vectors having a random set of nonzero
37. A. K. Noor and R. E. Fulton, Impact of CDC STAR-198 elements. Only the values of nonzero elements are stored along
computer on finite element systems. J. Struct. Div. ASCE with a secondary full vector called an order vector whose ele-
731-750 (1975). ments are all bits. The nonzero bits in the order vector indicate
38. A. K. Noor and S. I. Voigt, Hypermatrix scheme for finite the location of the nonzero values and zero bits indicate the
element systems on CDC STAR-100 computer. Comp. absence of values in the primary (value) vector. This capability is
Struct. 8(5/6) 287-2% (1975). still evolving and experience with its use is limited.

You might also like