AD-A009 261

PROBLEMS IN LAMINAR FLOW-TURBULENT FLOW

K. M. Case, et al
Stanford Research Institute

Prepa , for:
Defense Advanced Research Projects Agency

April 1975

DISTRIBUTED BY:

Knr
National Technical Information Service
U. S. DEPAHTMENT OF COMMERCE

afe*^

■riMBüMM H*.
 •^—^

135124

April 1975
Technical Report JSR-74-2

01
05
O
^ PROBLEMS IN IAMINAR FLOW-TURBULENT
«£ FLOW

By: K. M. CASE A. M. DESPAIN E. A. FRIEMAN

F. Vi(. PERKINS J, F. VESECKY

Conuact N... DAHC15-73-C-O370

AR PA Order No. 2504
Program Code No. 3K10
Ds ^ of Contract; 2 April 1973
Contract Expiration Date: 30 June 1975
D D C
Amount ot Contract $1,297,321

M/W 12 1975
Approved for public release; distribution unlimited.

iisisnTrEi
D
Sponsored by

DEFENSE ADVANCED RESEARCH PROJECTS AGENCY

ARPA ORDER NJ. 2504

77TVV

I Ä Ä
STANFORD RESEARCH INSTITUTE
Menlo Park, California 94025 ■ U.S.A.
Raproduead by
NATIONAL TECHNICAL
INFORMATION SERVICE
US Otpirtintnl o* Comm«fc«
„ A,
Copy No
8
* ,

/
Sptin»ti«l<(. VA. M1S1

The views and conclusions contained in this document are those of the authors and should not be
interpreted as necessarily representing the official policies, either expressed or implied, of the
Defense Advanced Research Projects Agency or the U.S. Government.

tf)

-t MBB
h-T^TTTrTZ^IF.CA-nON OF THIS PAGE (When Dg Entered)
READ INSTRUCTIONS
BEFORE COMPLETING FORM
REPORT DOCL'MENTATION PAGE
3 RECIPIENTS CATALOG NUMBER
2. GOVT ACCESSION NO
[I REPORT NUMBER
J3R-74-2 JJ2A^£3J2Al.
■ TYPE
5. „.- REPORT
'.TTT. OF occ^DT ii, PERIOD
PFRIOD cov(
COVERED
j 4. TITLE '.ard Subtitle)
Technical Report JSR-74-2
PROBLEMS IN LAMINAR FLOW-TURBULENT FIOW
6. PERFORMING ORG. REPORT NUMBER
SRI Project 3000
r? AUTHORi») 8 CONTRACT OR GRANT NUMBERS)
K. M. Case A. M, Despain E. A, Frieman Contract ÜAHC15-73-C-0370
F, W, Perkins J. F. Vesecky
'10 PROGRAM ELEMENT, PROJECT TASK
9. PERFORMING DHGANl/A I ION NAME ANU ADDRESS AREA 8i WORK UNIT NUMBERS

Stanford Research Institute

Menlo Park, California 94025 13. NO. OF PAGES
12 REPORT DATE

-11 CONTROLLING OFFICE NAME AND ADDRESS April 1975 92

Defense Advanced Research Projects Agency 15 SECURITY CLASS, (of this reponl

140C Wilson Boulevard U?:CLASSIFIED

14 %£» A^g^E ^BSS [H .H. .o. 5^5 OH.c.)'
15a. DECLASSIFICATION/DOWNGRADING
SCHEDULE

16. DISTRIBUTION STATEMENT (of this report)

Approved for public release; distribution unlimited,

D D C
r^
MAY 12 '975.
; IS3 ir> RIOCK •'0 if different from report)
17. DISTR'BUH' STATEMENT (of the ab.tr.ct entered m BIOCK O,
iEDTTEt
o—
IB. SUPPLEMENTARY NOTES

RDS (Continue on reverse side if n.ce,Mrv and ident.ty By o.ock numb«)

19. KEY WO
Transition
Turbulence
Numerical simulation

^TTSiTSACT (Continue on rev. d. ,f nece.-ry and ident.tv by COCK number) transition from
A program is suggested for the numerical simulation of the transition fro

The program is:

, A commitment for 4 to 5 years.
• , i n^nc: (at
Parallel calculations fat ursi;
first) bv
uy aat least two grouos. (continued)

F0RM
nn
LrU 1 JAN 73 1473 IFJriON OF THIS PAGE (When Data Entered)
FSiTION OF 1 NOV 66 IS OBSOLETE NATIONAL T:(.HNICAL
INFORMATION SERVICE
ggyäS r, äüäSJäS f'F THIS PA
^ (Wh 1
'" ^ ^2S^L
19 KEY WORDS (Continued)

20 ABSTRACT (Continued)

rao...tl« up o£ a .«11 advisory eroup (approximately 4 ne.bors)

* S acUvHorLrs In the llo.a who .oulä folio, the .orK closely
and provide direction.
. The expenditure of about $200,000 per year.

*« ho treated (in a prelir.ünary ordering) are:

The problems to be treateu vm " i

(1) Transition in the boundary layer over a flat plate.

(2) Transition in the wake of a flat plate.

(3) Transition in the flow over wedges (Falkrer-Skan profiles).

(4, Modificat.ons of the flat-plate transition ^'^^

roughness, compliant boundary conditions, and heaUng.

Furthe^ problems of interest that might eventually be treated are:

(5) Couette flow between rotating concentric cylinder...

(6) Poiseuille flow between parallel planes.

(7) Poiseuille flow in pipes.

In .ppee.1. » . eeeo-enaatloe 1. --^^ ^ T^V^Tr

for .eeuietloe el ' ^^ ~ ^.^l-Toiil compute,.. AppePUlces C

ffW\
ijW\
F0RM
JAN 73 ,1473
(BACK)
SEC
TjBITY CLASSIFICATION OF THIS PAGE (When Data Entered)
EDITION OF 1 NOV 65 IS OBSOLtTE KK

' '■"*

I ■ ■
 CONTEXTS

v
IlLLSTRATIONS

v11
TABLES

1
I INTRODUCTION

5
II THE FLAT PLATE
lD
A. Boundary Conditions
18
B. Calculations at Breakdown

19
III COMPUTER REQUIREMENTS

A. Introduction 19
20
B. Finite-Difference Method
20
C. Spectral Method
D. Comparison of Computing Time Estimates 25

29
IV WAKE OF A FLAT PLATE

31
V WEDGE FLOWS

33
VI POSSIBLE FUTURE PROBLEMS
33
A. Couette Flow

B. The Benard Problem

C. Poiseuille Flow
D. Modifications of the Flat-Plate Problem 36

VII ON CONTROLLING THE TRANSITION TO TURBULENCE 37

VIII THE IMPACT OF NUMERICAL SIMULATION OF TURBUIENCE 39

41
IX RECOMMENDATIONS

iii

*m
APPENDICES

A LINEAR-STABILITY ANALYSIS AS A PRACTICAL TOOL .... 13

B REMARKS ON A SPECIALIZED NAVIER-STOKES COMPUTER . 47

C A MODEL TO STUDY DOWNSTREAM BOUNDARY CONDITIONS 59

D BOUNDARY CONDITIONS
65

E THREE-DIMENSIONAL IN-PLACE FAST-FOURIER TRANSFORM

ARRAY
77
REFERENCES.
83

IV

/
 *a

ILLUSTRATIONS

Measured Velocity Distribution In the Laminar Boundary

Layer on a Flat Plate at Zero Incidence

Curve of Neutral Stability for the Wavelengths oö* of the

Disturbances In Terms of the Reynolds Number R for the
Boundary Layer on a Flat Plate at Zero Incidence (Blaslus
Profile
> 10
Curve of Neutral Stability for the Frequency p, of the
Disturbances and Wave Propagation Velocity c for the
Boundary Layer on a Flat Plate at Zero Incidence (Blaslus
Profile)
11
Variation of Amplitude of the u'-Fluctuation for Two
Neutral Disturbances In a Laminar Boundary Layer on Flat
Plate at Zero Incidence 12
Distribution of Intensity of u-Fluctuation Across Boundary
Layer: 145-c/s Wave, U /v = 3.1 X io5 ft-1 13
G Geometry of Flat-Plate Navler-Stokes Problem 19
7 The Form of the Neutral Curve In the (k,Re) Plane for a
Plane Poiseullle Flow 35
D-l Coordinate System 69
D-2 Stability Diagram for Boundary-Layer Flow. 73
E-l Perfect Shuffle Connections for 1-D FFT N = 16 81
E-2 1-by-4 Array of Cells for a 2-D FFT Network. 81

i
/

rtMi
 w mm "

TABLES

1 Three-Dimensional Finite-Difference Technique 21

2 Estimated Computing Time for Spectral and Finite-
Difference Methods 26
B-l Parameters of an Interesting Turbulence Calculation ... 50
B-2 Maximum-Performance Parameters for a Modern Single,
Pipelined and Cashe CPU Computer 51
B-3 ILMAC IV Parameters 52
B-4 Cellular-Array-Computer Parameters 56

vil

Jl — ^■M
 i INTRODUCTION

It is clear that a fundamenta] understanding of turbulence would be

of great practical significance. Since it is suspected that on the very

smallest scales turbulence has certain universal features, it might be

hoped that these would be revealed by a direct numerical simulation.

Unfortunately the problem of fully developed turbulence is one involving

very nu .y degrees of freedom. Indeed, if R is the Reynolds number, the

necessary degrees of freedom to be treated in a given dimension are of

3 4
the order of R In three dimensions this number is then of the order
9/4
of R (see, foi example, Landau and Lifshitz, 1959: for more detailed

estimates, see Case et al., 1973). Since time steps are restricted by
3
space steps, the number of needed calculations then grows as R (Case

et al., 1973). For the large values of R occurring in practice the

computations then become prohibitive with present (or soo.>-to-be-available)

computers.

Under these circumstances it is perhaps reasonable to back off a bit

and ask what can be done in the way of a numerical investigation of the

transition from laminar to turbulent flow, here, unless otherwii» indi

cated, we restrict ourselves for simplicity to the incompressible case—

i.e., our flows are described by the Navier-Stokes equations

2V 2
.1^: + v . Vv = -VP + vV v (1)
at ~ ~

V . v = 0 (2)

References are listed at the end of tho report.

M
 Knowledge of such transitions also has considerable significance for

practical designs. The hope is that we will be able to treat such prob-
lems with presently available computers, since:

(1) Transition frequently takes place at relatively low

Reynolds number.

(2) Transitions are frequently seen to develop throt-h a

number of stages in which the flow is relatively large-
scale (i.e., involving relatively fev. degrees of freedom)

Experimentally, the transition takes place in many different ways.

Indeed, for a given practical situation more than one mechanism may be

responsible. There are a number of ways in which one can characterize

the different possibilities. None of these is completely satisfactory

in that the descriptions are not necessarily unique or mutually exclusive.

One such cataloging is in terms of whether the transition occurs in a

flow that is inviscidly stable or not. Another rather convenient dis-

tinction is between transitions in "free" or "bound" boundary layers

(Sato and Kuriki, 1961). In essence, the difference is between flows

that become turbulent rapidly and those that become turbulent gradually.

It is the latter class that we expect to be most amenable to numerical

simulation. However, as we will see, the division is by no means sharp.

Accordingly, we consider .he calculation of transition in a variety

of simple situations that have shown different forms of transition. We

start with problems in which there is much detailed experimental infor-

mation available and some theoretical insight. As will be seen, this

knowledge is very useful in showing us what size calculation must be

performed and what one expects to be able to describe. Assuming that

the initial calculations are successful, we then indicate others that can

be done that hopefully will lead to methods, in which there is confidence,

for calculating transition in situations of practical interest.

/
*H
 Our program is then as follows. First we consider various situations

in which transition could be calculated. These are enumerated essentially

in the order of what we consider their priority. m each case we describe

the physics of the situation and our present understanding. In various

degrees of detail we discuss the limitations of the calculations (reso-

lution needed and uncertainties). An estimate is then made of the compu-

tation effort required. Second, we give some recommendations as to how

a program to do these things could be implemented. These include cost

per year, number of years, organization, and supervision.

It is felt certain that such a program would contribute significantly

to the present state of the art with a relatively modest expenditure of

funds. It is hop^d this will lead to a capability of doing some practi-

cal calculations.

—■—

■CM
 II THE FLAT PLATE

As a first calculation we suggest the numerical study of the transi-

tion in the boundary layer in the flow over a flat plate. The advantages
of this are:

(1) Codes ior this have been developed and some calculations
have been performed (Grosch, 1974; Orszag, 1974).

(2) Excellent detailed exptrimental information is available

(Klebanoff et al., 1961; Kovasznay et al., 1962).

(3) Considerable theoretical insight is available. Thus

the initial stage of laminar instability is well under-
stood (Schlichting, 1960). Theoretical models give a
qualitative picture of what happens in the later stages
(e.g., Benny, 1961; Stuart, 1962; Lin and Benny, 1962;
Greenspan and Benny, 1963).

There are also disadvantages:

(1) As indicated in our earlier report (cf. Schlichting,

1960, p. 386) there are some ambiguities concerning
downstream boundary conditions.

(2) At the last stages before complete turbulence is

obtained, small-scale motions are found. There is
some question as to whether a calculation designed to
follow the earlier stages of transition can resolve
these.

We return to these two points later.

To illustrate what is involved in the calculation of transition of

the boundary-layer flow over a flat plate, we give a brief (and idealized)

version of the experiment described by Klebanoff et al. (1961).

Over a flat plate located at x s 0, y = 0 there flows a fluid with

velocity Uoo in the x direction toward the plate. The pressure gradient

Preceding page blank

*■
is adjusted to be zero. A ribbon is located above the plate in a span-
wise (z) direction perpendicular to the flow at some distance downstream
from the leading edge. Under the ribbon there are strips uniform in
length and uniformly separated. With the ribbon staMonary the x component
(u) of the velocity is, fo.- any x and z, as shown in Figure 1.

FIGURE 1 MEASURED VELOCITY DISTRIBUTION IN THE LAMINAR BOUNDARY LAYER

ON A FLAT PLATE AT ZERO INCIDENCE. Source: Schlichting (I960).

The ribbon is then vibrated at a frequency f. At small amplitudes

of oscillation the results are as follows. For a fixed distance x down-
stream the velocity varies essentially sinusoidally in time and with wave-
length, which is the periodicity of the strips in the z direction. The
average velocity distribution as a function of y retains essentially the
form of Figure 1. As we go downstream the amplitude of the oscillations
in u first increases and then gradually decreases.
/

mtm j
 At a larger amplitude of oscillation the picture changes. As we
proceed downstream from the ribbon the situation Is first as above.
Further down, however, there develops * pronounced three-dimensional
structure characterized by large spanwl.e variations in wave amplitude
with "peaks" and "valleys" occupy ng fixed spanwlse positions. Associated
with this variation in amplitude there is also a spanwlse variation in
local mean velocity such that there is a defect at a peak and an excess
at a valley. Further downstream there occurs an abrupt increase in
amplitude at a peak that is characterized by a series of intense low-
velocity pulses-evidenced by "spikes" i„ an oscillogram of streamwise
fluctuating velocity. At first a single spike appears for each cycle of
the primary oscillation. The spikes increase in number as we go down-
stream and ultimately blend into fully developed turbulence.

Present theoretical understanding is the following. In the absence

of the vibrating ribbon the stationary laminar flow is adequately
described by the boundary-layer equations for the velocity [u(x.y).vCx,y),
w = 0]:

"öu vdu d u
dx V
dy " ~~2 (3a)
dy

du äv
dx dy (3b)

u=v=0aty=0 lim u = U

Those equations have the Blasius similarity solution

U00 v
u = U00 f(Tl) V =
(4)

\
 ^^"^

where

Ti = y.

With

f(Tl) = cp'CTl) i^CT]) = l/2[t\v'(T\) - cpl

we have

cpcp + 2cp =0 (5)

subject to

cp(0) = cp'CO) = 0 cp'(») = 1

Numerical integration then gives the velocity profile of Figure 1—in

excellent agreement with experiment.

Remarks;

(1) Two conventional measures of the boundary-layer

thickness used are:
(a) 6 = that value of y for which u = 0.9911^.
Numerically,

(6)
— F

(b) 6
••fKh—^
=

(2) For the experiment of Klebanoff et al. (1961) approxi-

mate numbers are U» M 1500 cm/s, v = 0.15 cm /s, and
x ~ 100 cm. TTien, 6 L.. ~ 0.5 cm. Typically
ribbon ribbon

M
 the variations described above occurred within 40 cm
of the ribbon. At this point 6 has changed by only
~20%.

(3) From Figure 1 and the above we conclude:

(a) The velocity distribution varies smoothly within

the boundary layer.

(b) After but a few 6 from the plate (^1 cm for the
experiment) the flow is essentially arbitrarily
close to the free-stream velocity.

Tc study the effect of external disturbances it Is natural to turn

to linear stability theory. Thus, in Eqs, (1) and (2) we substitute

u = u + 6u (7)
~ ~o ~

where u is the original laminar flow, and we retain only terms linear

in the perturbation 6u. Assuming that a time dependence 6u ~ e" ^ leads

to an eigenvalue problem. If it is found that Im üü = üü is > 0, then

initial perturbations can grow, and we have instability. In practice,

another approximation Is made. Stability locally is studied by assuming

that the profile (y dependence) is such that the one at a given position

Xo really extends from -oo < x < oo, (Since in the problems considered

here the true x variation of the profile is very slow, this is a readily

justifiable approximation). In this case th-? eigenvalue problem is

simplified. Since there is no explicit x-dependence we can limit our-

selves to investigating solutions with x and z dependence of the form

-(ox+ßz) ,
~ • . LSince Squire's theorem (Schlichting, 1960, p. 386) assures

us that the onset of instability occurs first when P = 0, the original

solutions of the eigenvalue problem were obtained for this case]. With

this form of x-dependence the eigenvalue problem can be regarded as one

in which a is given (real) and the complex uu is sought.

Figure 2 gives the results for an old calculation of the curve for

neutral stability in the 06 , R. = U 6Vv plane. It may be noted that

0♦ 00

*m _j
 1

K- y

FIGURE 2 CURVE OF NEUTRAL STABILITY FOR THE

WAVELENGTHS ah* OF THE DISTURBANCES
IN TERMS OF THE REYNOLDS NUMBER R
FOR THE BOUNDARY LAYER ON A FLAT
PLATE AT ZERO INCIDENCE (Blaslus Profile).
Source: Schlichtmg (I960).

for instability, R6+ must be greater than a critical value R ~ 500

and Qr6* must be less than about 0.36 (i.e., there is a minimum wavelength
X 2n6 /0,36 17 56 G6 f
min " " - * - or instability). For the experiment
described above this wavelength is X ~ 3 cm, while the distance x
min c
downstream from the leading edge where instability can possibly occur is
x ~ 80 cm.
c

For the experiment, however, it is somewhat more relevant to con-

sider a real ■ and find the, in general, complex o. (With a = Im a < 0,
we have spatial amplification.) A neutral curve in the u), R
plane is
6*
shown in Figure 3. We see that there is a maximum frequency for which
such spatial amplification holds. This is

tt> 6
max
* 0.19

10

^m tt*
 ^ -*»

FIGURE 3 CURVE OF NEUTRAL STABILITY FOR THE

FREQUENCY ^ OF THE DISTURBANCES"
AND WAVE PROPAGATION VELOCITY c,
FOR THE BOUNDARY LAYER ON A FLAT
PLATE AT ZERO INCIDENCE (Blasius Profile)
Source: Schlichting (I960).

For the experiment described, this is:

t ~ 300 c/s
max

We note also from Fieure •? thot +»,„ u

ipure 3 that the phase velocity (c = ./») of the
ni.nt ..^1 . . _ r
neutral waves Is of order 0.4U or less.

«.e .„eoretioal picture of the early st.ee, of the e.per^ut

descrfhe. .. theu the follo.lng. *. osclUatlng ^ ^^ _

of .avele^th X > , «. Wese are theu a.pllf^. as „ go do.ustrea..

the lumal perturhattou U s„fflclently SM11 „ ^ ^

«he rlBht-ha,d „eutral curve hefore the a.p,„tuaes are large e„„ugh to
»of. the lluear assunptfou. As m eo tarther d„„stream ^ ^^^^^

-iU theu he dampea out. Ou the other haut,, if the iui.i.i dlstllr|)ante

.. lar6e euough it .ill he anplifi«, „„„!„ .he „„stähle regiou aowu-

stream to a point .hero liuear-stabilitv theorv „„ ,
unity theory uo luuser applies aud
cannot be used for prediction

11
mm
1

One further result of linear-stability theory of importance for us

i.s the shape of the eigenfunctions corresponding to the unstable mode.
Fron. Figure 4 we see that they vary rapidly in the part of the boundary
layer between y = 0 and y = 0.2 - 0.46. (For the experiment, this number
is y - 0.1 - 0.2 cm.) Outside this region the variation is rather smooth

16 ■ —^

N
1 1

HP
i» A ■;
nmyScMchhng
MemnrnmrScfiubam
and SkramstM
1 1
0.8
1
Q.S

a* f+ 1
|

02 %
0 MI —' •
-02
Yl
ir *^-
-0*
0 0* cu» h? ;/f ;> 7 ?> . ^/I

I
FIGURE 4 VARIATION OF AMPLITUDE OF THE u'-FLUCTUATION FOR
TWO NEUTRAL DISTURBANCES IN A LAMINAR BOUNDARY
LAYER ON FLAT PLATE AT ZERO INCIDENCE. The curves
labeled I and 11 correspond to the two neutral disturbances I and
I. m Figure 2. Source: Schlichting (1960).

The further development of the flow toward turbulence is somewhat

less know, theoretically. From models (Benny, 1961; Stuart. 1962; Lin
and Benny. 1962; Ore .nspun. 1962; Greenspan and Benny. 1963) and experi-
ment (Klebanoff et al., 1961) it appears that the amplified two-dimensional
Tollmien-Schlichting wave interacts, as a result of nonlinear terms,
with true 3-dlmensional disturbances produced by the spacers, mis
combination then gives rise to the distorted boundary layer which in
turn results in the production of the high-frequency disturbances. While
[
12

*m J
 1

the primary contribution of a numerical simulation would be to allow the

development of the nonlinear interaction, two qualitative conclusions
can be drawn:

(1) Resolution in the spanwise (z) direction must be

sufficient so that the expected peaxs and valleys
can be adequately differentiated.
(2) While the Tollmien-Schlichting waves have their
largest amplitudes and sharpest variation close to
the plate (~0.l6) the fluctuations near breakdown
have their largest values much farther out (~0.56)—
cf. Figure 5. From this we conclude that even at
these distances the resolution cannot be too coarse.

FIGURE 5 DISTRIBUTION OF INTENSITY OF u-FLUCTUATION ACROSS BOUNDARY

LAYER: 145-c/s WAVE, U^/v = 3.1 X 105 ft-1. Source: Klebanoff et al.
(1961).

Let us now turn to the numerical simulation of the above experiment,

Cltarly, finite computer capacity restricts us to a discussion of the
fljw to some finite region of space and time. In the x (strepmwise)
direction this should begin somewhat downstream from the leading edge of

13

w*—ä
 the plate ,„ assure the lnm.1 establish.»,« „t the laM„,r no. hut
upstream ,„. the rlhhou to dln,ln.sh spurious .tt.ct. due to incorrect
boundary conditions. A reasonable .uess .„uid he to st.rt at a p<sltl„n
<a 5>
- ^. "»st""" '«. the ribbon (.here X^ Is the n,l„lmu. „„stahle
•.oUMen-Scbllchtln, .a.elengtb,. ^„strea« tro. the ribbon .o „uld
lik. to include all positions tin the onset o. turbumnce. EXperlTOnt.ny

this is ^X^. „„.ever, as Indicated, later ambiguities In the boundary

conditions suggest that the end position be chose.-, as far denstrea. as
feasible. A compromise might he that the region «n * to be described
is of the order o. (10-20)^. since In the sp,„wise (., dlrectlon „
are describing a periodic behavior, .e can restrict ourselves to a dis-
tance edual ,„ one »avelength of tb, spat.lse perturbation, in the
direction perpendicular to tbe plate .e can, since the laminar How and
... Perturbations go to zero rapidly outside the boundary-layer thlchneas
6, restrict ourselves to a distance ol a ,„ , <say, 26). ,„ „^ „
•ould llke to lollo» the lie. lor some slgnllicant „umber ol oscin.tlons
tt T he the period of the ribbon the time ol Interest may be „1 the order
of (5-10)T.

No. what do these dimensions Imply about the number o, calculations

to be perlormed, The problem Is to solve Bqs. «, and », in the region
described, subject to suitable boundary conditions (discussed belo.,
To estimate, .e Imagine the calculation done by llnlte-dmerence methods
l. dl dimensions, „f spectral methods are used, a rule ol thumb might
• -
be tha. the number ot modes In a given direction be -,l/2 , 1/3, „ the
number ol grid points. ,. the x.dlrt.ctlon .e .^ ^ ^^ _ ^^

of the order ISX^. Assuming that this is adeauately resolved by 8

points per .avelength .e have southing l.ke 160 points in tbe .-direction-
I.e., in the notation belo., L - 160. m tbe z-dlrectlon .e have a
distance of only one perturbation .avelength, but to have a chance of
describing tbe variation near breakden we need a fine resolution. A

14

/
guess is 16 points in this direction (N = 16). In the 6-direction we
have a total distance 26, but as we have seen in the region 0 i y < 0.26,
there are very rapid variations. Hence, with finite-difference method»
one would use either some coordinate mapping or a varying grid size.
Assuming ehe latter, one might use of the order of 20 points for the
first 0.36 and 20 irore for the remainder—i .e. , of the order of 40 grid
points. The sizes of the time steps are essentially determined by
stability requirements—i.e.,

vAt < Ar

where v is some appropriate velocity and Lr is some mesh spacing. Since

U is the largest velocity and it is in the x-direction, a reasonable
00

criterion (Grosch, 1974) might be

With the numbers suggested above, this is

0.375 -4
At < • 2.5 X 10
~ 3
1.5 X 10

-3
For T = 1/150 c/s = 6.7 X io~ , this implies of the order of 30 time
steps per cycle and ♦hexefore a number of time steps Nt ~ 300 to 500.

Computer requirements based on these numbers are discussed below,

First, however, we include a few caveats.

A. Boundary Conditions

A well posed problem for solutions of Eqs. (1) and (2) would be
one in which the initial velocity is prescribed everywhere within the
region described and the velocity is given everywhere on the boundary

15

^M*
 for all time (subject to j^n • u ds = 0). Unfortunately we do not know

this Information. At y = 0 we do. There all components of the velocity

are zero. At y = 26 it is clearly correct to a very high degree of ap-

proximation to take the velocity as that of the free stream. At the

upstream edge ol our region it is reas-jiiafcle to assume that the velocity

is of the Blasius form plus whatever p-M turbation one might want to put

in. In the spanwise direction a periodicity requirement is reasonable.

What we do not know is the velocity at t) e downstream edge of our region.

Two somewhat ad hoc suggestions have been made. Grosch (1974) suggests

"extrapolating" from the velocity inside. If L is the downstream edge

of the region, this formally is of the fovm

L
du
ua.y.z.t) - j u(x,y,z,t) g(x) dx ♦ |H (x.y.z.t) h(x]
0 dx
I Ok
0

In some sense this may be regarded as a kind of radiation condition,

which might indeed be reasonable. How well the actual choice made of

g(x) and h(x) approximates such a condition is not clear. Also, there

is an uncomfortable situation in that the pressure at x = L is calculated

from bu/bt, while to advance forward in time du/dt is calculated using

the resulting p. The possibility of an instability occurs.

Orszag (1974) proceeds by dropping certain terms as small. With the

emasculated equations he demonstrates uniqueness if at all boum'ary

points either P or y . n is given as well as y for points such that

0 (Note: this
~ * "out * ' demonstration involves an additional, but
reasonable, approximation.) The difficulties here are:

(1) We do not know p or y . n on the downstream boundary.

(2) We do not know y so that we cannot even determine when

v . n < 0.
~out
(3) The approximations involved need to be Justified. (In
fairness it must be noted that Orszag has recently

16

/
MM M
 suggested s. me modified boundary conditions about which
we have insafficient information to permit making an
evaluation.)

We would lixe to suggest that results obtained will be ratl.or in-

sensitive to the choice of downstream boundary conditions. There are two
arguments:

(1) For Tollmlen-Schlichting waves the group velocity (c )

is of order l/3Uro. Here then cg ~ 500 cm/s. Then to
travel a distance of the order of 50 cm from the ribbon
to the downstream edge requires ~ 1/10 s, which is
approximately 15 periods of oscillations, and one might
not even follow the developments much further in time.
(2) The essential use of the velocity on the boundary is to
compute the pressure within our region. This can be
obtained by taking a special solution of the Poisson
equation and adding to it a solution of Laplace's
equation

V P = 0

chosen to satisfy the boundary condition. If the y

variation of the boundary condition is ~ eiky, then
the fall-off in the x direction is:

? ~ P e '
o

Now a reasonable guess for the y variation is k26 ~ 2rt.

. (Remark: We are thinking of deviations from the Blasius
profile, /or the latter, of course, P is independent
of y.) This suggests, then, that the boundary conditions
imposed at the downstream edge will have little effect
a few wavelengths upstream. It is for this reason that
we have suggested that the downstream edge of the calcu-
lation region be some distance beyond where we hope to
find a turbulent transition.

A further argument that suggests things are insensitive to the down-

stream boundary conditions is found by noting tl at the velocities (phase

and group) of the Tollmein-Schlichting waves are usually small (~ 1/3)

/
17

■rt^
compared to the free-stream velocity. This suggests that any such
disturbance will be convected downstream and have little effect upstream.

In Appendix C we give a simple model for which the arguments for

exponential fall-off and downsitream convection are verified in detail.

However, it is obvious that none of the above arguments are rigorous

It is essential that whatever ooundary conditions are used should be
varied (within reasonable limits) to see how far upstream we need to be
in order to have confidence that the results are not artifacts caused by
the assumptions made.

B. Calculations at Breukdown

As we have noted, there comes a poirt where the boundary layer is so

modified by the nonlinear wave interactions that it becomes strong?y un-

stable. At this point high-frequency oscillations suddenly appear. In

model calculations the instabilities grow by orders of magnitude in a

fraction (~ 1/10) of the primary period. The scheme described above will

not be able to follow such rapid variations. At best, a large irregular

change in velocities between two successive time steps might be found.

However, even this may not occur—the large space and time differences

may average out such irregularities. One approach to proceeding further

is suggested by the fact that breakdown (Klebanoff et al., 1961) appears

to occur at very localized points. Hence, we can ;.top the calculation

when at some x,z there is a profile that according to linear stability

theory is highly unstable. Then we consider only a small region in the

vicinity of this point. For this region we introduce a much smaller

space-time mesh and then integrate the Navier-Stokes equation with initial

conditions obtained from the previous calculation.

18

MM
 Ill COMPUTER REQUIREMENTS

A. Introduction

The basic problem under consideration is that of flow over a flat

plate as illustrated in Figvre 6. L, M, and N are the numbers of mesh

V (NORMAL TO PLATE),
M= 32-64

FLOW
x (FLOW OIRECTIONI,
L- 128-256

FLAT PLATE OF INFINITE

SPAN IN (x, zl PLANE

z (SPAN DIRECTION), N- 8-16

FIGURE 6 GEOMETRY OF FLAT-PLATE NAVIER-STOKES PROBLEM

points in the x, y, and z directions at which one calculates the flow

variables v and P. Two approaches to the solution of the Navier-Stokes

I equations on digital computers are consideied here: the finite-difference

scheme of Dr. Chester Grosch, and the spectral method of Dr. Steve Orszag.
We have examined the finite-difference method In considerable detail
with the help of Dr. Grosch, but the spectral method has proved more
difficult to investigate, as noted below. The objective of this section
is to make some computer-run-time estimates for both methods on three- /
dimensional problems of interest. The results should not be taken as a

19

X »». rt*
comparison of the two methods (as we discuss below), but rather as an

indication of the computer run time and therefore cost that one can

expect on optimized "production runs" for useful prrblems.

B. Finite-Difference Method

The finite-difference method for flow over a flat plate in two di-

mensions (x,y) is discussed in detail by Grosch (1974). The method is

fairly easily generalized to three dimensions. The basic changes are

2
that the Poisson equation for the pressure V P-Q = 0 must be solved in

three dimensions and that three components of v are required, thus

complicating the finite-difference equations. The Poisson equation is

solved by taking a two-dimensional Fourier transform in the (x,z) plane

for each of the M values of y in the mesh. Me then solve (LN) tri-

diagonal equation systems—i.e., one system for each mesh point in the

(x,z) plane. This solution yields the Fourier components of P for each

plane perpendicular to the y axis. An inverse Fourier transfrom then

yields P. A set of three finite-difference equations yields the three

components of v and the loop is completed. Table 1 summarizes the

operations required and the total number of operations per time step.

Multiplying the terms in Table 1 by their appropriate computation times

yields the total compute: time (AT ) per calculation time step:
fd

AT =« LMN it (95 + 3 log LN) + t (81 + 2 log LN) + 28t I (8 )

fd ( a 2 m 2 sf

where t , t , and t are the particular computer's add, multiply, and

am s
memory access times.

C. Spectral Method

It is impossible to determine exactly how Orszag implemented his (

calculation. If his report (Orszag, 1974) is examined it will be

20
 (0
u
>> V
0 X

11 M m 0) x
M

M
s :
5
c
0
C
U
c
+
X
•H
w X
c
n
w
u

w
E ■ ■0
N

-
—
Q
I O CO

A
B o +
II in
5 <y
05
«
i
■c
CO
a.
N
l-l
CO c
bfi o
c
■H a
I o
pH
>
t §
E
E
o
PH
3
>i

■
IB U —
u
e
o
1 in
c
1 §
0)
3
?
§ 0
in
C "—•
••-)
rH
■
U
0

in
I
■~ •H 0 10) •H U
+J •(-> «i
r-t a Ü 0) U

I a I
^ s ^s h '—' e e
3 4) S SH
2 4) o
u
O ■H 1—1 0 0 •H C
l-l t, ••-1 b£ >-^ h fl 0) E in o
eg
U l
K cx
•rt
•o
1
0)
T3
0
a a m <H
<H
■H
Q
U
e
o
u
!fl rH>
0) 1 +J 0 -H Q J
E a •H Q ^ I 0) a
3
t/l
tn
M
«H
0 E >.
n
0
«1
1
N -M
0
-
S
0) ~ u
0)
a
CJ

0) I o
0)
u
E
3

N 5

21
noticed that no detailed discussion is given about the solution technique,

Verbal conversations with Orszag helped some, but it is difficult to

judge the validity and efficiency of the calculations. In at least one

case, the report contains conflicting statements concerning the step-

integration method employed. '"Adams-Bashforth" is indicated on p. 31,

but "Runge-Kutta" is used in the actual code. Since the computer code

still contains "bugs," it is not clear just how valuable the results of

the study were. Although it is not evident from the report, it is abso-

lutely true that Orszag has made significant contributions to the field.

The general method can be described as a part spectral method and

part finite-difference method. According to Orszag, 20 two-dimensional

real FFTs (fast Fourier transforms) and two one-dimensional FFTs are

required for each plane normal ,o the flow per time step, and the time

spent in FFTs is 60% of the total computer run time. Also, he stated

that it took 9 seconds per time step for L = 128, M = 8, and N = 64.

Thus, a general formula for estimating the run time on a CDC-7600 can
be derived as follows.

First, let us consider FFT operations:

(1) Complex transform in one dimension requires

log (N) butterflies with

(i)
•1 multiplies + 2 additions/rotation + 4 additions/butterfly

= multiplies + 6 additions for each butterfly yielding

3N log (N) additions and

2N log. (N) multiplications in all.

(2) Real transform in one dimension requires

(f)h (5) + 1 butterflies or

22
 log N multiplications and
N log,
(1)-]-"
(?) log N additions.

(3; 2-D complex transform (M X N) requires

M[- log (N) butterflies (1-D Transforms)

2 2

followed by

butterflies (1-D Transforms)

N - log M
2 2

NM log M + log N] = NM log2 MN butterflies

2

or
|4 multiplication
NM log2 (MN) X |6 addition

(4 MN log2 (MN) multiplications

=
je MN log2 (MN) addition.

(4) 2-D real transform (M X N) requires

l 4 log.
(5) + i butterflies followed by

l0g (MUi butterflies

Al 2
■ *! h (") •*. (i) * 1 - (^ ^ "^ + 2 io ,,,, rflies

or
MN log« MN multiplies

— log„ MN additions

Orszag has estimated that the (20L) two-dimensional real FfTs and
(St) one-dimensional FFTs require about 60% of his computing time. For
the 7600. multiplication is about twice the time of an addition. Define
one operation as an addition; hence, multiplication is equivalent to two

operations and a butterfly is 14 operations: '

23 i

^^m^
 ^^■•-

l0e M +
2L (ID FFTs) m 14L
[(?) 2 (5) l0SN operations

20L (2D FFTs) « 20 X 14L |^j log^MN) operatl ous

The total run time (ATg) per computation time step on the 7600 Is then

AT
(0.6/
(14L) 5 MN log2(MN) ♦ (5j log2M . (Hj ,^1
AT •K'L [ll7 MN log2(MN) 4. 5.8 M log M + 5.8 N log
'I
where K' Is a calibration factor (equal to the add xlme In seconds).

Now, from Orszag's experience on the 7600 we know that AT « 9s for L =

s
128, M = 8, and N = 64. Substituting this into the above equation yields
K' « 1.3 X 10 , so the effective add time (t ) is about one us.

Now, for a single-calculation time stop we will allow 60% of the

run time to the 20L two-dimensional FFTs-l.e., neglecting the two one-

dimensional FFTs, we allow 30% of the run time to the solution of tri-

diagonal equation systems, and 10% of the run time to i4 memory accesses

per mesh point. So from Orszag's experience we can derive an approxi-

mate formula giving A*, the computing time required for a single time
step in the Navier-Stokes calculation:

AT, • LMNta (45 log MN) + LMNt (30 log^MN) + 14 LMNt

(9)

where L, M, and N are the numbers of mesh points in the x (flow),

y (normal), and z (span) directions; and t , t , and t are the cc
am s
effective addition, multiplication, and memory access times.

24

*m
D. Comparison of Computing Time Estimates

The computing-time estimates (per time step) for the finite difference

and spectral methods discussed above are given by Eqs. (8) and (9). The

particular computer's effective add, multiply, and memory access times

are denoted by t , t , and t . Two features are common to both methods:

am s
(1) The computing time increases approximately linearly
with the number of mesh points.

(2) The multiplication terms are more heavily weighted

than the addition terms when one considers that
multiply times are typically about twice as long as
add times.

However, there are also some interesting differences. The finite-

difference method is more sensitive to memory access time and therefore

will fare relatively worse or. a memory-limited machine. For a typical

application the spectral method is most heavily governed by the multi-

plication time, while the finite-difference method is about equally de-

pendent on the multiplication and memory access times.

The computing-time estimates given below should be qualil.led on

several points. First, the estimaces are intended only as a very rough

gui.de to the sort of computing times one might expect on rather highly
optimized "production runs." We have used effective values of t^, t^,

and t derived from the experience of Dr. Steve Orszag on the CDC 7600
s
computer where he has taken care to reduce the run time by optimization

and the use of many assembly-language-coded subroutines. The finite-

difference method of . Chester Grosch has never been optimized for the

7600, so the figures below are in a sense a rough projection of what the

finite-difference method could do if optimized in a similar fashion.

Spectral methods are thought to give better resolution of the hydrodynamic

variables for a given mesh size. Thus, for calculations of a given

accuracy, the number of mesh points required could be less for the spectral

25

■ " "■'

ri^
 •m^^

method. Finally, the algorithm for the finite-difference method has been

examined in more detail than has that for the spectral method. Given

these uncertainties, one should regard the estimates below as only rough

indications of the range of computing time required, and not as a com-

parison of the two methods.

For problems of interest the numbers of mesh points required in the

x, y, and z directions fall in the following ranges: L ~ 128-25Ö, M -

32-64, and N - 8-16. In Table 2 we have estimated the run times for

maximum (~ 250,000 mesh points), minimum (~ 33,000), and typical (~ 66.000)

problems. AT, the computer time per calculation time step is calculated

from Eqs. (8) and (9) using "effective" values of t ~ 130 ns, t - 260 ns
am'
and ts - 1 us. These effective values are derived from the experience

of Dr. Steve Orszag on the CDC 7600 computer, but we have used them for

estimating both AT^ and AT^. Typically, 300 to 500 time steps are re-

quired to carry a calculation to the transition stage, so we have taken

T = 400 AT as the total run time.

Table 2

ESTIMATED COMPUTING TIME FOR SPECTRAL

AND FINITE-DIFFERENCE METHODS

Problem L AT AT T T
M N LMN fd s fd s

Maximum 256 64 16 262,144 20 s 40 s 2.0 hr 4.0 hr

Minimum 128 32 8 32,768 2 s 4 s 0.2 hr 0.5 hr
Typical 128 64 8 65,536 5 s 8 s 0.5 hr 0.9 hr

Note: L = Along flow; M ^ Normal to flow; N = Along spsn.

26

^MM
 ^

Again we should emphasize that these figures should not be taken as

a comparison of the fInlte-dlfference and spectral methods-they are too

rough ne main ,olnt here Is that for suitably optimized codes the run

times are not unacceptably lor*. Given a typical cost of ~ $700 per hour

lor a 7600. the production run cost for the "typlcl" problem Is about

$500 Optimization and assembly language coding (assumed here), which

take advantage of the architecture of a given computer, can give a time

reduction as large as a factor of five. So one could expect » unopti-

runs."

27

' , IIIIPIW»

^m -
 mm^

IV WAKE OF A FLAT PLATE

A prob:em closely related to the above transition in the boundary

layer of a flat plate is that o.'' transition in the wake of a flat plate.

Some reasons for considering this problem next are:

(1) The calculations are very similar to the previous one.

A suitable program for the flat-plate calculation
would require modest changes in order to treat the wake.

(2) Detailed experimental information is available for

comparison (for example, see Sato and Kuriki, 1961).

(3) The nature of the transition is rather gentle. In

contrast to many other turbulent transitions, this
one is not achieved by the development of sharp
bursts. Accordingly, it may be expected that
transition can be followed numerier tly very far.

Drawbacks to this calculation are:

(1) As previously, the downstream boundary conditions

are ambiguous.
(2) The applications to practical design are somewhat
indirect,

Our understanding of this transition runs as follows: The Goldstein

(laminar) wake (Goldstein, 1933) is unstable according to two-dimensional

linear stability theory. Thus a small perturbation is amplified when

one proceeds downstream. Eventually nonlinear effects take over. (The

growth rate is not the predicted exponential, and harmonics of the in-

duced perturbations appear.) Further downstream a distinct three-

dimensional pattern appears and the flow becomes more and more irregular.

However, we emphasize that no sharp bursts or spikes appear. This should

be an ideal problem for numerical simulation.

/
Preceding page blank

ri^ ^i J
 ^

Since, as Indicated, the problem is so much like the flat-plate

transition, computing requirements will also be pretty much the same,

30

/
MM
 ^^ wmm

V WEDGE FLOWS

TTie numerical simulation of transition in wedge flows would be of

considerable interest. The reasons for this are:

(1) The calculation is very similar to the flat plate.

[Indeed, one code (Grosch, 1974) has been written
to include this possibility.] The essential reason
for this is that the boundary-layer equations for
the flat plate are Just a special case of that for
the wedge. Thus, if ßn is the included angle, the
potential flow is

■
U(x) = U x (10)
o

wnere

m = ß/{2 - S)

and the boundary layer is described by the equa'lons

(Schlichting, 1960, p. 143):

m + 1 " m^ I m - 1 ,
u = U x f'(T])
o ■ "M—^— vU x f + f'
V 2 o m + 1 '
(11)

U
. m + n1 o m -
M = y A\— x (12)
V 2 v 2

in /x2.
+ ff" + ß[l - (f') = 0 (13)

subject to f(0) = f'(0) = 0 f'^) = 1

We note that with ß = 0 this is just our flat-plate
equation. (The factor-of-2 difference occurring
between Eqs. (5) and (13) is due to a different
normalization.)

31

mtm mi^s^m
 (2) By varying the parameter ß we can study the effects of
pressure gradient. Thus, with ß > 0 we have accelerated
flow (and therefore greater stability), while for ß < 0
we have deceleration (and more instability).

(3) With ß = 1/2 we have the equations of the boundary layer

for a rotationally symmetric flow.
(4) Measurements of the boundary-layer transition on a flat
plate with mild, favorable pressure gradients (DeMetz,
1974) are available for comparison with the calculations.

As we have indicated, the computing requirements are as for the

flat-plate case.

I
32

^m
 VI POSSIBLE FUTURE PROBLEMS

There seem to be many different mechanisms for transition to turbu-

» lence. For an understanding of these mechanisms, it would be desirable

to do numerical simulation in relatively pure situations. Here we give

a list of some of these and an indication of why they are of interest.

(Since the calculations are of a somewhat different nature than those

described above and have less direct practical application, we think

they should be done at some later date. Accordingly, we have not made

any computing estimates. However, it seems that they are problems of a

magnitude similar to that of the problems discussed above—but computa-

tion would probably require new codes.)

A. Couette Flow

This flow between concentric rotating cylinders is, of course,

classic. Extremely good experimental results are available (Donnelly,

1963; Coles, 1965). From a theoretical standpoint this is a very clean

problem. The ambiguity of the boundary conditions disappears. In the

radial direction we have well defined boundary conditions on the cylinders.

Azimuthally we have periodicity. In the longitudinal direction periodicity

is a reasonable requirement.

When the inner cylinder rotates more rapidly, there is seen experi-

mentally a rich set of phenomena (Coles, 1965). When the rotation rates

are changed we pass from the original laminar state through a multitude

of different states with well organized How patterns. In addition,

hysteresis effects manifest themselves- i.e., the state achieved depends

on the past history. It would be fascinating to reproduce these effects

by numerical simulation and it should be possible to do so.

33

^M«*ii
 Lin's calculation
Lock's calculation

ffe,/J

FIGURE 7 THE FORM OF THE NEUTRAL CURVE IN THE (k.Re)

PLANE FOR A PLANE POISEUILLE FLOW. Source.
Monin and Yaglom (1971).

2. In Pipes

This is a particularly interesting situation since, in contrast

to all other problems we have discussed, the flow is apparently stable
with respect to infinitesimal disturbances. We say apparently since:

(1) Extensive theoretical efforts using linear-

stability theory have failed to find any
unstable modes.
(2) Experimentally the transition Reynolds' number
seems to go ever higher, the lower the ambient
disturbance level. (Of course, neither of these
is a proof.)

At least two ideas as to how the observed turbulence originates

have been suggested. One is that there is a finite amplitude instability.
If this is the case, numerical simulation seems to be of little value.
The observed transition in terms of turbulent bursts suggests that in a
calculation we would either be in a laminar region (of little interest),
or in a turbulent region where we cannot resolve the flow. The second
suggestion (Smith, 1960) is that turbulence originates in the inlet
region before the full Poiseuille profile develops. If this picture is
correct, the numerical calculations would be similar to those for the

35
 flat-plate boundaryy layer.
laver T+ -«-—- . L„„
„ should po8slble to calculate ^
the m, to transition to mil, „evo^pod turbu.onoo.

D
- """"""tlons g the Fl.t.p,... o„..,.n

M~U, suocoss^u, co.pletlon of «, flat.plate calculatlons ,

--.' - ^»«U, nn.ntc: wlMtl .^ „ „^ ^

"*"" Very ,mta ^ -'"-t.on. ^3, the effects of suction

™ugnne8s. a„d complUnt bou„darles ^^ ^ treated ^ TO^ ^^ ■

o ar co <,i,ion8 imposed the

: :r : "
coma a.so no tno.teo. ■ **- *• •«-• - •» --.
An ,ddltlonal eq„atlon (tor ^ temperatur

•ou. ^. to be „so., an. tho aPproPrlat8 va.a.ton tn .soostt, „eeds

* .ncludod. WHUe so.e reCodl„g .ould ,0 necessary, ^ ^
".asnitudo of tn, eo^utat.on noeded .ould not ho chan.od.

36

m^^^
J J
 VII ON CONTROLLING THE TRANSITION TO TURBULENCE

The development of a three-dimensional numerical simulation code

will give us a powerful new tool for assessing methods to control the

transition to turbulence on a wide range of bodies—i.e., not just bodies

of revolution and two-dimensional airfoils. We use the phrase "controlling

the transition" in a general sense so that studies of effects that sub-

stantially advance the transition (e.g., excessive surface roughness)

are also included in the scope of the investigations.

Let us first turn to ways to calculate and control linear Tollmein-

Schlichting Instabilities. Here a number of methods have already been

suggested: (1) compliant surfaces, (2) changing the equation at state

(e.g., adding heat, polymers, etc.), (3) suction, and (4) control of the

pressure gradient. Except for pressure gradients, which have enjoyed

wide use, the optimum ways to use Methods 1 through 3 remain to be devised.

Furthermore, even though one is dealing with linear-stability problems,

there are important nonlinear effects in the ways the laminar boundary

layer and potential flow patterns are altered by the measures used to

delay the transition. For example, adding heat changes the viscosity,

and the viscosity gradient in turn alters the laminar boundary layer in

a direction toward stability. Clearly, heat added at one location on the

surface of a body will diffuse across the boundary layer and have a much

smaller effect downstream. Thus the combination of heat transport via

diffusion and convection, coupled with the linear-instability mechanisms,

becomes a complicated problem that is probably most easily solved by a

direct simulation rather than linear-stability analysis. The present codes

must be generalized to include heat transport, of course.

37

-—

MM
 The second-class methods to alter the transition involve nonlinear
changes in the flow within the boundary layer—e.g., nonparallel flows,
surface roughness, and abrupt changes in the body shape. As examples
we can cite the early transition to turbulence that occurs on some rough,
blunt bodies and the use of small span-wise or flow-wise grooves to
stabilize the Tollmein-Schlichtlng waves. Also, swept-wing airfoils
have boundary layers with nonparallel flow. The numerical simulation
approach gives us a single technique that can handle this wide variety
of problems.

Overall, our studies indicate that numerical simulation techniques

have sufficient resolution to accurately predict transition and that they
can be adopted to circumstances where the laminar boundary layers differ
appreciably from the simple parallel-flow ones now commonly used.
Initial studies of such phenomena would take a year to complete, and a
comprehensive program would require three to five years.

/
38

*M J
 VIII THE IMPACT OF NUMERICAL SIMULATION OF TURBULENCE

Let us suppose that three-dimensional numerical simulation techniques

have successfully reproduced many of the phenomena that occur in flat-

plate flows. What comprises the next generation of problems both in the

practical and research areas?

An important class of practical problems concerns the simulation of

the transition to turbulence in configurations where linear-stability

analysis fails or is excessively complicated due to geometry. These

problems will require the development of a more flexible code that can

handle both fairly general boundary conditions (e.g., surfaces rough on

the scale of the boundary layer) and heat transport. Fortunately, almost

every body can be represented by a sequence of simulations each of which

represents a small portion of the surface.

The impact of the numerical simulation of turbulence is that It will

be a single technique that will handle many problems not only of a linear-

stability type but also inherently nonlinear, such as finite-amplitude

surface roughness. The ability to compute the combined effects of both

linear and nonlinear processes for initiating transition is a strength

unique to the numerical simulation approach.

The next generation of research problems should be chosen to help

us understand how well the simulation code represents fully developed

turbulence. Estimates clearly indicate that the inertial range will be

described poorly if at all. For example, high-frequency bursts of

turbulence may not occur in the correct frequency range. On the other

hand, many of the overall properties of a turbulent boundary layer (such

as drag, heat transport, etc.) may be quite well represented since they

39

\,
depend much more on the larger, energy-containing eddies. Therefore,
there exists a strong possibility that numerical bimulation may produce
a useful enough representation of the turbulent boundary layer so that
the effect of body shape, etc., on the transport properties of the layer
can be calculated—at least in a semi-quantitative manner. Such a pro-
gram would require codes with flexible boundary conditions and carefully
thought out sub-grid closure schemes. Its impact would be to produce a
first-principles calculation of the arbitrary coefficients that occur in
turbulent-boundary-ltyer theory, and, more importantly, to see if there
are circumstances in which these coefficients vary by a significant
amount.

/
40

«rtM
 IX RECOMMENDATIONS

A program of numerical simulation that could be expected to have a

significant impact on our knowledge of, and ability to predict, the

transition from laminar to turbulent flow might consist of the following:

(1) A commitment to the program for a significant length of

time. Crash programs seem to be counterproductive. Four
to five years is a reasonable time scale.
(2) An aspect of competition. There should be at least
two groups performing the calculations. In the beginning
they should be calculating as closely as possible the
same problem, and with the same parameters used in the
experiments. Later, when confidence has been obtained
the groups could begin dividing up the problems suggested.
It is reasonable that the groups use somewhat different
computation schemes. A number of such schemes suggest
themselves. It would be efficient to use the program
to evaluate their relative merits. The schemes used to
date~i.e., those of Grosch and Orszag—seem sufficiently
representative.
(3) A small advisory panel. This might consist of a group
(~ 4) of active workers ^.n the field. The members might
be an experimentalist, a theoretical hydrodynamicist, a
computer expert, and a numerical analyst. They would
be expected to work closely with the computing groups.
Fopefully, they would spot difficulties or uncertainties
in given calculations and direct the program (i.e., while
we have given a tentative ordering of problems to be
done, it would be expected that the advisors would con-
tinually reassess the next stages of the program in the
light of what had been accomplished).
(4) An expenditure of the order of $200,000 per year. This
mighi produce between 30 and 100 runs on one of the
problems outlined above. We envisage a few man-years
of effort and computation on a computer comparable to
the 7600. This sum may or may not include support for
experimental work that might be found advisable.

41

M
 Appendix A

LINEAR-STABILITY ANALYSIS AS A PRACTICAL TOOL

Preceding page blank

43

MM^M -
 Appendix A

LINEAR-STABILITY ANALYSIS AS A PRACTICAL TOOL

On a properly designed and fabricated body, the transition from

laminar to turbulent boundary layers occurs through linear Instabilities
whereby free-stream fluctuations In velocity, pressure, and temperature
enter the boundary layer and are spatially amplified via the Tollmeln-
Schllchtlng waves. A sufficiently large amplification results In non-
linear effects that Initiate the transition to turbulence. The transition
prediction method of A.M.O. Smith and his colleagues Is the only predic-
tion method that now makes use of the spatial amplification concept~a
process that must occur physically. Indeed, Smith's e method has been
the most successful predictor In ARPA's program as well as In predicting

transition for the various airfoils.

But Smith's method has not received wide use because the spatial
amplification rates are derived from the solution to an eigenvalue problem
involving 4th-order differential equations. The solution to these equa-
tions, which must be done by computer for an arbitrary laminar boundary
layer, are conceptually straightforward but quite laborious In practict.
A well documented and widely available computer program is called for.

We therefore recommend that ARPA fund an effort to produce a computer

program (most likely in TORTRAN) that will run on a number of the more
generally available scientific computers. The core of the program will
be the 4th-order eigenvalue solver. Auxiliary parts will Integrate the
spatial amplification factor along the body, and calculate the laminar

boundary layer, potential flow, etc.

Preceding page blank

^MMM
 The wide availability of such a program would allow body designers
to test proposed bodies against Smith's criterion on an in-house basis,
hopefully allowing a more rapid convergence to optimum designs.

The need for this computer program is an example of a situation in

which an increased understanding of the basic physics has led to more
sophisticated design criteria that can be implemented only by a computer
program. Of course, some generality is lost from the older, more gross
engineering criteria. But, it is only through such a program that the
effects of the details of body shape on the transition to turbulence can
be quickly evaluated.

In closing this section, we should remind the reader that transition

can occur through channels other than linear Tollmein-Scnlichtir.g ampli-
fication. High levels of free-stream turbulence, excessive surface
roughness, and poor body fabrication are examples. The spatial amplifi-
cation method often does not. work in such cases. But it does have a good
record of success in those projects where designers have taken the
trouble to eliminate these other sources of turbulence.

46

WA
 Appendix B

REMARKS ON A SPECIALIZED NAVIER-STOKES COMPUTER

47

^H^MB
 Appendix B

REMARKS ON A SPECIALIZED NAVIER-STOKES COMPUTER

1. Introduction

Reference 2 has considered the question of what can be accomplished

with modern machines such as the ILLIAC IV, IBM-370/195, and CDC-7600.

The purpose of this appendix Is to consider the wisdom of designing a

special computer that could greatly exceed the capabilities of the current

machines In solving only a specialized class of problems—1 o,, the

Navler-Stokes equations, and similar partial differential equations

arising in plasma physics, hydrodynamics, weather prediction, geology,

etc.

a. An Interesting Calculation

An interesting numerical simulation well beyond the capabilities

of current machines is the fully resolved three-dimensional turbulent

4
fluid flow for a Reynolds number of 10 , This requires a grid of
.4 9/4 9
(10 ) =10 mesh points for each point in time. Each mesh point must

represent the velocity and pressure fields so that three velocities and

one pressure term are required at least for the present Instant in time

and the immediate past. The past values are needed in order to mirch
Q
forward in time with a numerical integration procedure. Thus 8 X 10
scalar variables are required, and about (10 ) in 10 == 101 arithmetic
operations are required. A more detailed count of arithmetic operat-ons
of an efficient finite differencing scheme for 10 grid points Indicated
that 5 X io additions and 3 X 10 multiplications are needed ^er time
12
step. A rough guess ID that about 10 instructions would be required

49
Preceding page blank

am Mn^^^^m
 12
to implement these operations so that about 2 X 10 read accesses from

storage are required for each simulated time step. These parameters are

summarized in Table B-l.

Table B-l

PARAMETERS OF AN INTERESTING TURBULENCE CALCULATION

Three-Dimensional Geometry

Reynolds number: R = 10^

Resolution: 109 spatial mesh points per time step

Memory size: 10 sealer variables

Arithmetic operations
(finite-difference method)

Additions: 5 X 10 per time step

Multiplications: 3 X 10 per time step

Instruction execution rate: 12 per time step

10iÄ

Memory bandwidth: 2 X 1012 read accesses per time step

b. Current General Purpose Computers

Table B-2 lists some critical state-of-the-art parameters of

modern computers. Note that with a simple array system of secondary-

storage access systems, more than 2 X 1010 X (2 X IQ )" =10 s would

be required Just to update each grid point for each simulated time step.

Parallel access to the secondary storage will reduce this time, however.

The instruction execution time for a very-high-performance single CPU

Q 9
(control processing unit) is at best 25 X 10' s, so that 25 X 10 X
2 X 1012 = 5 X 10 s or about 14 hours per simulated time step would be
9
required. Thus it seems infeasible to simulate 10 grid points since /

50

*mm
 Table ß-2

MAXIMUM-PERFORMANCE PARAMETERS
FOR A IODERN SINGLE, PIPELINED AND CACHE CPU COMPUTER

Direct-access secondary-storage capacity: ,-.11 bits maximum --i.e.,

10 I
about 2 X ICr variables

Direct-access secondary-storage bi >dwidth: 107 bits/s maximum--i.e.,

about 2 X 106 variables/s

Arithmetic operation time: 25 X 10"9 s minimum (pipe-

lined)

Cache memory access time: 25 X 10~9 s minimum

many time steps would be required in a practical problem. Clearly, the

usual von Neumann computer architecture cannot be used, and other choices

should be considered.

c, ILLIAC IV

The ILLIAC IV (Bell and Newell, 1&71) is a radical departure

from classical computer architf -.tur^. It was designed for array-type

calculations. The critical parameters of the ILLIAC IV are given in

ble B-3. The most limiting facto, is the limited secondary storage of

only 109 bits as compared to the 10 bits that are required. Even If

the ILLIAC should be suprlemented with large disc drives, the bandwidth
9 10
to secondary storage would be less than 10 bits/s as compared to 60 X 10

bits needed for one simulated time step,. It would take more thar 60 X

10'9 = 600 S just to read In data fo- one simulated time step. Calcula-

tions would require about 2 X 10^ X 2 X 240 X lo"9 + 64 = 15,000 s per

s<T.ulated time step, if all input-output and grid-point data interchanges

are ignored. Thus even an extensively enhanced ILLIAC IV cannuL oe us

/

M
 mmm

Table B-3

ILLIAC IV PARAMETERS

64 processing elements

Add time (each element): 240 ns

Multiply time (each element); 400 ns
2048-64 words of 240-ns
memory/element

Secondary storage capacity: 109 bits

Secondary storage bandwidth; 10 bits/s (to all elements in parallel)

to attack this problem, and other types of computer structures need to

be examined.

2. Computer Architecture for Array Calculations

a. Solution Methods

Nonlinear partial differential equations can be solved by a

variety of techniques such as finite difference, spectral, psuedospectr.. 1,

and Monte Carlo methods. A special computer designed for any one of these

methods will not be optimal for any of the others. Also, while it appea-s

that finite-difference methods are the most flexible relative to bounJary

complexity, etc., the pseudospectral methods are much better when regular

surface boundaries exist. Furthermore, integration algorithms are in a

dynamic state of development and better algorithms aru quite likely to

emerge in the near future. Thus, only very flexible and general computer
structures should be considered.

52
 ^^■"

b. Computing Requirements

What are the common characteristics of a large class of non-

linear partial differential equation solution methods? First of all,

frr very arbitrary boundary conditions some finite-difference and suc-

cessive over-rexaxation methods must be employed, riowever, if well

behaved boundaries (planes, spheres, cylinders, etc.) exist, then the

pseudospectral methods offer large savings in computing effort. In fact

9
it appears that the 10 grid problem will be attacked in the foreseeable

future only by pseudospectral or other methods that ,ake advantage of

particular features of a given problem. Finite-difference and relaxation

methods require only that information be passed between neighboring grid

points. The spectral techniques on the other hand need global informa-

tion at every grid point in order to calculate a Fourier or other trans-

form (Orszag, 1974). However, a very large class of transforms can be

implemented through the fast Fourier transform (FFT). Thus the archi-

tecture of the computer should be such that grid points can interchange

information with their close neighbors and also be able to perform 3-

dimensional FFTs on the quantities of the grid points.

F-om Table B-l it is evident that to solve the Navier-Stokes

equations, for a large number of grid points, about ten variables per

grid and 1000 operations per grid point are required for each simulated

time step. More complicated problems involving compressible fluids,

electromagnetic flells, etc., might double both ol the quantities so that

in designing a relati 'ely flexible computer, perhaps 32 variables per grid

point should be selected and 2000 to 4000 operations per grid point per

time step assumed. If it is desired that a time step be simulated in

about a second, then about 3200/32 = 100 operations per second are needed

for every variable.

53

^m
 c. Performance Tradeoffs

Very fast pipelined logic (a high-performance processor) could

perform about 2 X IQ7 operations per second while a slow serial processor
3
might only be able to do about 3 X 10 operations per second.

For one second of execution time per simulated time step,

2 X 10/2000 2: 10 grid points could be processed by a single high-

4
performance processor with (10 to 32) X 10 registers per processor

(i.e., approximate 10 bytes). However, an interconnected array of

109 X 10 or 10 X 10 high-performance processors would be required to

handle the 109 grid-point problems. This approach seems to be far beyond

near-term technology hence is therefore considered infeasible. Even if

the calculation time were increased considerably and the logic speeds

could be greatly improved, an impossibly large array of very-high-

performance computers would still be required. In any case, the pro-

cessors of such an array would be very similar in structure tc the present

high-performance machines discussed earlier so that beyond creating

special FFT hardware attachments to the usual sort of pipelined processor

(or array processors for ILLIAC IV), it appears that little can be gained
by a special computer architecture using high-performance operations.

An interesting situation develops, if however, the low-

performance serial organization is considered. If one slow serial pro-

cessor were "devoted" to one grid cell and hence 32 variables, the re-

sulting computer structure would be a three-dimensional array of slow

serial processors, each with about 32 registers, connections to its

neighbors (six each), and connections to implement an PFT (see Appendix

E for duration of the FFT) in three dimensions (six additional connections

per cell).

A more efficient arrangenent in terms of total number of inter-

3
connections is to use one processor for every (N ) grid points; for the

54 /

riMM J
lowest-speed serial processor, this could require as long as ten seconds

to simulate one time step. This may very well be an optimal arrangement
for a special-purpose computer devoted to solving hydrodynamic equations.

As N is increased, the arrangement of cells is nut changed but a few

more instructions per cell are required for individual grid-point selec-

tion within a given cell. Table 3-4 illustrates the possibilities for

several values cf N. It is significant to note that by 1980, it may be

feasible to construct a computer to solve the 10 grid problem for about

the cost of the ILLIAC IV (in 1970 dollars).

3. Conclusions

It can be concluded that:

(1) Dircc'- numerical simulation of turbulence for most

important and interesting problems cannot be accomplished
with present-day computers or even with general-purpose
computers proposed for the near fu'.u'-e. Hence, many of
these simulations will be impossible without siecial-
purpose computers that have been custom designed to
solve partial differential equations.
(2) Algorithms for solving partial differential equations
are in a dynamic state of development. Thus any pro-
posed special-purpose computer must be very flexible
relative to solution method as well as to the kinds of
problems it can handle. In particular, it should be
capable of finite-difference calculations and relaxation
calculations as well as spectral and pseudo spectral
calculations for sets oF partial differential equations
that arise in hydrodynamics, plasma physics, aerodynamics,
global weather, etc. Hence, each grid point must be
able to exchange information with neighboring grid
points and must be capable of in-place, three-dimensional
fast Fourier transformation of the grid-point quantities.

(3) A special purpose very-high-performance computer could

be organized for the 109-grid-polnt problem by using a
large-variety of high-performance, expensive components.
However, the many different types of components required
and their diverse interconnections would not allow any
I
55

mrtm J
 1

1 1 <H
0
2
1

ta 3
ij
2
a
■
10
B
O
- Ä 1

in
a 2 8
■H
*-•
0)
■o

ration and
L y •H
I
i
eg
i

hnology
0
10 -

0
0
i
«->
i
<H
1
x:
a
■o
Z

•o
IH
0

• ? 0 u 0
B 2
1974 state- § 2CO
Ü (1 in ai B
B +J -H •H
a! Q (i

11
W
no logy 8 N
0
«1
0)
<H
0

1 Q
i §
Is »
0)
■H
+J
U
o)
0>
rH
-H
*>

<H -
0 0 -o o o
co (H U
<0
-.
o
to
o
ir>
o
•o pH
1 u in u X
1 0> o X
1» X
o o X X X (0

1s
CM
fH in o
1 1 in CO CM (0 rH

u 0 E- .M
CO

■
s■
IA B rH 00
■a
CHE
c cj (0
in
CM
(0 CO CM
CM
§
0 4-> i-l CM CO m t* o
11 o U 3 Ei rt
rH o 0 u
a 1 m o u
to
& w a
in

J
i <H B
0 O PH
•H »H
Number
Connect
per Ce

i eg CO
CM
00
CM
o
CO
i-H
CO
CM
CO
CO
eo
M
CO

<H
0 in i-t
RAM Bit
per Cel

.
Number

| O o
8
CM
§
eo
1
rH

<H m

■in
0 h rn
0 i-l
o
CO §
o
§ | | § §
O
gj

E
in

bo
O

h
§ o 8
l-( CM Tf (0 o (fi o"
1 S SL CM R
k
in o

o B in
in (0 eo • CM 0> CO
•H o> i-( 00 CM rH
fH CM o
h o J Iß CJ CO m C»

^^ 2
5 -a a.
II H II II II II n n II n
CO CO
3 H IN
CO
CO
CO CO CO eo eo n
10 (0 oo O) o
o | rH

56

«■
 ^^w

regular construction pattern, so that the complexity

would be overwhelmingly great. Such a machine appears
impractical for the near future (the next 10 to 15 years).

(4) A cellular-array computer, composed of a single inexpensive

general-purpose minicomputer driving a very large,
regular array of identical cells of a single type, each
cell of which represents one or more grid points, does
o
appear to be feasible for the 10 -grid-point problem
by the early 1980s.

(5) A cellular array of more modest size might be very

effective for calculating some specialized problems
in the transition to turbulence, plasma instability,
etc. hence, some smaller array may be worthwhile prior
to the 1980s, when 1SF technological advancements
will probably allow the construction of the large
array for a cost that is not totally unreasonable.

(6) A very modest investigation of the cellular-array

computer appears to be justified, since many
interesting questions must be answered before such
a machine could be constructed.

(7) Extensive simulation of the array machine should

precede any commitment to hardware. This could be
accomplished by a continuing research program at
modest cost.

(8) The potential benefits of a powerful computer of the

type proposed hero seem to be very great. Some
further investigation is certainly indicated.

57

mum J
 1

Appendix C

A MODEL TO STUDY DOWNSTREAM BOUNDARY CONDITIONS

Preceding page blank 59

" —' imii«n«»i

- MMfc g
 Appendix C

A MODEL TO STUDY DOWNSTREAM BOUNDARY CONDITIONS

In the main text a number of arguments were given to suggest that

downstream boundary conditions cause little effect at a few-boundary-

layer thickness upstream. Here we give a simple model In which we can

verify the statements.

We consider Couette flow between a stationary plate at y = 0 and a

plate moving with velocity U in the x-directlon at position y = 6. The

laminar solution of the Navler-Stokes equations is

v = u (y). 0, 0 where u (y) = Uy/6

o o

(Thus we are approximating the Blasius-type profile by a straight line.)

Further, we neglect viscosity and restrict ourselves to flow linearized

around the laminar solution. We wish to see how a disturbance at a point

affects the flow upstream and downstream from that point. Writing

v = (u + u , v , 0) we have the equations

~ oil

a«,"1 a» 1 au
' o
+ u + v
at o OX i ay ay

av au
—i + u — = -apyay
at o ax i

ax ay

Introducing a stream function cp by

Preceding page blank 61

■M
 u = *£ ,,--*■
i öy i ax

we readily find that

/a_ + U
u S_\ (iLi +h ^2) , 0
\ öt o ax/ \dx2 ^yz/ "

(Note thnt ä2cp/ax2 + äVöy2 is just the vorticity of the perturbed flow.)

Taking ordinary Fourier transforms with respect to space 1nd one-sided

ones in time—i.e.,

iojt
cp(x,y,(i;) = I e cp(x,y,t) dt ,

and

cp(k,y,uü) =
1 e
ikx
cp(x,y,u)) dx

--we see that

2»
d cp 2 « _ W(k,y,0)
05 -
(C-l)
2 i(ku -w)
dy o

where W(k,y,0) is the spatial Fourier transform of the initial value of

the vorticity.

Explicitly,

W(k,y,0) =1
1 e
-ikx
Acp{x,y,0) dk

If we introduce the Green's function Gd.y^k) by

62

JL
 sin ky sin k(6 - y)
G(y,y/,k) = - y s y'
k sin k 6

sin ky' sin k(6 - y)

y a y
k sin k 6

the solution of Eq. 'C-l) Is given by:

cp(k,y,(ü) = G(y
üly y
y, K;k) nKy',o) y
,
' • l(ku (y')-^)
/ o

Inverting the time Fourier transform gives

f1 -ihu (y')t
©(k.y.t) = I e GCy.y', k) W(k,y'pO) dy'

and then

lk[x-u (y')t]
cpCx.y.t) = dy' |
./'../ e G(y,y',k) W(k,y'0) dk

0
(C-2)

Now, let us consider the effect of an Initial disturbance at x = 0,

y = y . Thus,
o

Acp(x,y,0) = 6(x) 6(y - y )

Then W(k,y',0) = 6(y - y ) and Eq. (C-2) becomes

o

lk[x-u (y )t]
V(x,y,t) = e G(y,y ,k) dk
2n J o

63

^^MMBMM
 n

Appendix D

BOUNDARY CONDITIONS

"

/
65

*m
 I ■ I

Appendix D

BOUNDARY CONDITIONS

There is a general class of problems, of which the transition problem

is an example, that require a careful examination of boundary conditions.
We first examine the question of what must be specified in order to arrive
at a unique solution, and we then examine a set of boundary conditions
for the specific problem under consideration. Lastly, we show that in
principle there is a computational scheme that follows from this speci-
fication.

The Navier-Stokes equations for an incompressible fluid governing

two flows v and v are:
1 2

^1.8 0 2
-Tr+Zlt2 • ^1,2 ^ -VP1,2 +
^1,2

V • v =0
-1,2

Defining

= V
^ ^1 " 2

P = P - P
1 2

We find

^ 2
it * 2l * ^1 " i ' 7*2 = -VI + vy v

Preceding page blank 67

—mm
 7 • v = 0

* ä ~ä ~ä ^x ~ ~ ~2

Thus,

öv
rr + v • yv + v • yv = -yp + w v
n i- ~i ~ ~ ~^ ~

V • v = 0

The energy in the difference flow,

■/i'
dt

is governed by

ÖE (I v2 irr
ät j v ' ~i r+ -^ • v
^2 * y)dt ■ -Jv • py + vj 7 • vv. v

which may be rewritten

äv av
as. _ f/
i (v • yv . v + v
i i\
at

- | dslPn .v-vn.7^-+n v —|

If we suppose that vi = ^ on the boundaries, the surface terms vanish.

Assume that 7v2 is bounded and its maximum is gi' en by

M = max(vv )

SB

■ I^^I^
Then

p ^ 2M(t)E
ot

We can then write

2/ M(t')dt'
0 ^ E(t) ^ E(0)e

from which it follows that if E(0) = 0, E(t) also vanishes.

We have therefore proved tiiat specifying the velocity on the bound-

aries is enough to provide a unique solution.

We next turn to examine the specification of the boundary conditions

appropriate to the problem at band. The coordinate system is shown in

Figure D-l. The steady flow satisfies the exact equations

öv dv ^2 V
. 0 A 2
O V
X z d? / X X
v -— + v —
x dx z dz
dx dz

ÖV
I
dv
2
h2
+ V
x dx Z ÖZ

BOUNDARY
LAYER
V

/
/ SPACE

/
y/ S FLOW

FIGURE D-1 COORDINATE SYSTEM

- MM

MM
 dv äv
X z
ox dz

The usual physical picture of boundary-layer theory involves the assump-

tion that the flow is more slowly varying in the x direction than in the
z direction. Thus we may write

— = E — € « 1
dx äex

and find

dv dv ,ov „ d v
.e+v .^_€+v/ ^e2 + _x
x dcx z öz äex j 2 2
d(ex) äz

2 2
dv äv A
O V A
O V
ÖP Z z
v -— e+v -— = - -— + v e +
z o€x z dz az 2 2
d(€x) hz.

öv äv
X z
r— € + -— = 0
dex öz

The divergence equation clearly requires

v = ev
z z

which yields

dv dv / d2v , d''v \
x —
v -— + V V € +
x dex r dz dex 1 i/ 2 e 2 /
\d(ex) dz /

2 2
dv —
dv ^
o - v dA "
v
z z dP 2_ + v z 1 z
v -— + v r— e +
x dex z dz dz 2 2 e 2
e ^ s
^(ex) ^
dz

70

^^^M^^l^
 In order to balance the viscous and inertial forces we then must have

v = v«

which finally yields the standard boundary-layer equations, which we re-

write in the original variables

dP

dv dv a2v
x _x dP X
+
x ox z äz ' dx 2~
dz

öv dv ?> v
» _Z z
x ix z dz 2~
dz

hv dv
X _z
öx äz

Estimating the terms yields

dv d2v
x äx x 2
A
oz

or

v ~ —
x 2
z

z z

This immediately suggests the Blasius similarity solution

71

_i
 v ■ Uf(Tl)
x

with

Tl = z/Jvx/U

v =-g(1fl)
z z

This leads to the nonlinear equation

f f " + 2f "' = 0

and the boundary conditions

f = 0 f' = 0 z = C

f ' = 1 z -» a»

Since these solutions are only approximate, part of thr^ flow lield that

develops in time in the computation will result from the equation trying

to predict the actual solutions. We look for solutions of the form

-igt ißy ir>(x)dx

e e
f = f(z)e

and we know that the resulting equation is a form of the Orr-Sommerfeld

equation which is fourth-order in z. Two boundary conditions are provided

by the vanishing of the perturbed flow at the bottom and top, and two more

involving the derivatives are provided by the vanishing of other components

of the velocity. The resulting eigenvalue equation

72 /

riM i^fe J
 h(uj,ß,a,R) ■ 0

patial growth. In general, defining the boundary-layer

yields the local s

thickness

6' -Jvx/U

and the Reynolds number

U6'
R6 =

Xds curves of th. I.« given in Flgnre 0-.. The n.tu.l experiment of
yie

GIVEN FLOW, X INCREASING

D-2 STABILITY DIAGRAM FOR BOUNDARY-LAYER FLOW

FIGURE

Klebanoff et al. (1961) corresponds to point A for the upstream condition.

To save calculational time, the numerical calculations may be started at
point B. which is 100 cm downstream, and the calculation can be stopped

at point C.
A su^ry of a physically reasonable set of boundary conditions that

at least is mathematically unique la then giv. n by

(1) Global constraint, / dS v . n •■ 0.

(2) Bottom plate, v = 0, no slip condition.

/
73
 (3) Top of region, vx = UC»), free stream.
(4) Span-vise, the flow is symmetric. Th-is, vy = 0 at the
sides, and öv /M and övz/öy both vanish. Although the
original uniqueness proof was carried out for specified
v, the surface terms vanish except for

J ds n VVT =Jdx dz ^x ^ + vy ^ + vz ~)

which vanishes for the specification of derivatives

given above.
(5) Upstream. As discussed above, the conditions are

v = U(z) + 6v
x x

v = 6v
y y

v = 6v
z z

where U(z) is the lUasius profile corresponding to a

position 100 cm downstream from the entrance, and 6v
is a linear combination of twr- and three-dimensional
Tollmien-Schlichting waves. These waves are chosen
to correspond to the driving frequency of the ribbon
in the Klebanoff experiment and to have the periodicity
in y imposed by the spacers.

(6) Downstream. Here we assume tbit the conditions imposed

in (5) are simply carried downstream to a location at
x = L corresponding to point C in the stability diagram.
We know that the spatial growth is converted downstream
with a velocity c that is less then U^, Thus, for some
period of time, chosen so that transition occurs in the
volume of the calculational domain, the downstream
boundary conditions will not change.

We now wish to show that, in principle, the specification given above

yields a method of numerical calculation. At t = 0, the velocity xield

is specifiel everywhere and for all t, v is specified on the boundary.

/
74

^M
From the Navier-Stokes equations and V • v = 0, it follows that

V P = -V (v • Vvj

A Greens function G satisfying

V G = -6(x - x')

yields the solution

= I ds (Gn • VP - Pn • VG) - I GV • (v • Vv) ds

/

Since P is unknown on the boundary, choose n • VG = 0 there. We are

guaranteed then that G exists. From the normal component of the Navier-

Stokes equation on the surface

— v-n + v'Vv'ns -n • VP + vV v • n
at ~ ~ ~ ~ ~

we see that n • VP is determined in terms of known quantifies. Thus,

2
ff
P=lGvVv*n-— v.n - v • Vv • n ds - J G(V . v • Vv) dt

and is determined everywhere. Now, knowing P we use

dv
— = -VP + vV v - v • Vv
dt » «. ~

and the knowledge of v everywhere to advance v forward in time at every

interior point of the calculational mesh. In a crude sense this represenis

a proof of the existence of solutions to these equations.

75
 Appendix E

THREE-DIMENS'ONAL IN-PLACE FAST-FOURIER TRANSFORM ARRAY

Preceding page blank 77

*,^m -
 Appendix F

THREE DIMENSIONAL IN-PLACE FAST-FOURIER-TRANSFORM ARRAY

Assume that it is desired that an array of omlln is to be permanently

connected with as few connections as possible such that three-dimensional,

in-place, fast Fourier transforms can be made of variables stored within
the cells, mat is, if the elements of matrix A, a k are stored in

cellijk, then matrix B with elements b^ are to be calculated according

to"

B = F[A]

L-l M-l N-l

%=£EZ i=0 m=0 n=0

a
*m eXp
\*n /i£
[2n \L +
im
M +
kn\]
X )\

Stone (1970) has shown how a one-dimensional FFT on an array of cells can

be accomplished in place by a "perfect shuffle" interconnection between

cells. Only two inputs and two outputs per cell are required. Despain

(1974) has shown now FFT operations such as those used within the cells

used by Stone can be implemented without calculating or storing the

trigonometric coefficients usually employed (cordic methods). This makes

practical an arr. - of simple identical cells that can be made to perform

an in-place transform of variables stored in the cells.

Three-dimensional transforms are generally accomplished as three

successive groups of one-dimensional transforms. However, it is possible

to simultaneously transform in all three dimensions at the same time.

This is of great advantage for the three-dimensional array of cells, since

the "perfect shuffle" connections are permanently connected in all three

Preceding page blank 79

^m
dimensions and parallel data transfer (and calculations too if desired)

can proceed simultaneously and at a much greater rate than if separate

transforms are taken in succession.

The connection pattt n is symmatric in all three dimensions. Figures

E-l and E-2 illustrate these patterns for one and two dimensions.

(
80

mtma^mm J
FIGURE E-1 PERFECT SHUFFLE CONNECTIONS
FOF 1-D FFT, N = 16

FIGURE E-2 4-BY-4 ARRAY OF CELLS FOR A

2-D FFT NETWORK

81

^M^^MM
 REFERENCES

Bell, C. G. i.id A. Newell, Computer Structures: Readings and Examples

(McGraw-Hill Book Co., Inc., N.Y., N.Y., 1971).

Benny, D. J., "A Nonlinear Theory for Oscillations in a Parallel Flow,"

J. Fluid Mech., Vol. 10, No. 2, p. 209 (1961).

Case, K. M. , F. J. Dyson, E. A. Frieman, C. E. Grosch, and F. W. Perkins,

"Numerical Simulation of Turbalence," Technical Report JSR-73-3,
Stanford Research Institute, Menlo Park, Calif. WiZ).

Coles, D., Transition in Circular Couette Flow," J. Fluid Mech., Vol. 21,
p. 385 (1965).

DeMetz, F., "Low Speed Boundary-Layer Transition Workshop," July 16, 17,
1974, The Rauu Corporation, Santa Monica, Calif.

Despain, A. M. , "Fourier Transform Computers Using CORDIC Iterations,"

IEEE Trans. on Computers, Vol. C-23, No. 10, pp. 993-1001 (October
1974).

Donnelly, R. J., "Experimental Determination of Stability Limits," in

Proc. Symp. Appl. Math. , Vol. 13, pp. 41-53 (1963).

Goldstein, S., Proc. Roy. Sos. A, Vol. 142, p, 545 (1933).

Greenspan, H. P. and D. J. Uenny, "On Shear-Layer Instability, Breakdown,

and Transition," J. Fluid Mech., Vol. 15, No. 1, p. 133 (1963).

Grosch, C. E., "Predicting Transition from Laminar to Turbulent Boundary

liyer Flow," Report, Old Dominion University Research Foundation,
Old Dominion University, No iolk, Va. (1974).

Klebanoff, P. S., K. D. Tidstrom, and L. M. Sargent, "The Three Dimensional

Nature of Boun-^ry-Layer Instability," Fluid Mech., Vol. 12, p. 1
(1961).

Preceding page blank

83

Ml
Kovasznay, L.S.G., H. Komoda, and B, R. Vasudeva, "Detailed Flow Field
In Transition," in Proceedings of the 1962 Heat Transfer and Fluid
Mechanics Institute. F. E. Ehlers, J. J. Kauzlarich, C. A. Sleicher,
Jr., and E, Street, eds. (Stanford U. Press, Stanford, Calif., 1962).

Landau, L. D. and E. M. Lifshitz, Fluid Mechanics, p. 123 (Addison-Wesley

Publishing Co,, Reading, Mass., 1959),

Lin, C. C. and D. J, Benny, "On the Instability of Shear Flows," Proc.

Symp, Appl, Math,, Vol. 13, No. 1 (1962).

Monin, A. S. and A. M. Yaglo.n, Statistical Fluid Mechanics. Vol. 1, p. 130

(M.I.T. Press, Cambridge, Mass., 1971).

Orszag, S. A,, "Fourier Series on Spheres," Monthly Weather Review.

Vol. 102, No. 1, pp. 56-75 (January 19747! '

Orszag, S, A., "Numerical Simulation of Boundary Layer Transition on a

Flat Plate," Research Report No. 33, Flow Research Inc., Seattle,
Wash, (May 1974).

Sato, H. and K. Kuriki , "The Mechanism of Transition in the Wake of a

Thin Flat Plate Placed Parallel to a Uniform Flow," J. Fluid Mech.
Vol. 11, p. 321 (1961). "

Schlichting, H,, Boundary Layer Theory. Chapt. XVI (McGraw-Hill Book Co.,
Inc., N.Y., N.Y,, 1960).

Smith, A.M.O., "Remarks on Transition in a Round Tube," J. Fluid Mech,.

Vol. 7, p. 565 (1960). '

Stone, H, S., "Parallel Processing with the Perfect Shuffle," IEEE Trans,
on Computers. Vol, C-20, No, 2, pp. 153-161 (1971).

Stuart, J. T., "On Three-Dimensional Non-Linear Effects in the Stability

of Parallel Flow," Advances Aerona t, Sei., Vol. 3 (Pergamon Press,
Oxford, 1962).

84
/

tfM