Lecture 22

MA471 Fall 2003

 Advection Equation
• Recall the 2D advection equation:
C x C y C
a a 0
t x y

• We will use a Runge-Kutta time integrator

and spectral representation in space.
 Periodic Data
• Let’s assume we are given N values of a
function f at N data points on the unit
interval.

• For instance N=8 on a unit interval:

0 1/8 2/8 3/8 4/8 5/8 6/8 7/8 8/8

 Discrete Fourier Transform
• We can seek a trigonometric interpolation
of a function f as a linear combination of N
(even) trigonometric functions:
k N / 2
If  x   
k  N / 2  1
ˆf eik 2 x
k

• Such that:
kN / 2
f  x j   If  x j    ˆf eik 2 x j for j  1,..., N
k
k  N / 2  1

j
where x j 
N
 Transform
• The interpolation formula defines a linear
system for the unknown fhat coefficients:
k N / 2
fj  
k  N / 2  1
ˆf eik 2 x j for j  1,..., N
k

jN
ˆf  1 N N
fe
 ik 2 x j
k j for k      1,...,
N j 1 2 2
Or:

f  ifft fˆ 
fˆ  fft  f 
Code for
the DFT
Code for
Inverse
DFT
 Fast Fourier Transform
• See handout
 Spectral Derivative
• We can differentiate the interpolant by:
k N / 2
If  x   
k  N / 2  1
ˆf ei 2 kx
k

k N / 2
dIf
 x     ik 2  f k e
ˆ ik 2 x

dx k  N / 2  1
 Detail
• We note that the derivative of the k=(N/2) mode
 iN2 2 x ˆ 
e fN 
 2 

iN  iN
2 x
ˆ 
• is technically: 2  e 2
fN 
2  2 

• However, as we show on the next slide – this

mode has turning points at all the data points.
 Real Component of N/2 Mode

i.e. the slope of the k=(N/2) mode is zero

at all the 8 points..
 Modified Derivative Formula
• So we can create an alternative symmetric
derivative formula:
k  N / 2  1
dIf 1
dx
 x 
2

k  N / 2  1
 ik 2  f k e
ˆ ik 2 x
Spectral Differentiation Scheme
1) Use an fft to compute:
ˆf for k    N   1,   N   2,..., 1,0,1,...,  N 
k      
2 2 2
2) Differentiate in spectral space. i.e.
compute:
 N N
 i 2 k  f k for k    2   1,...,  2   1
ˆ
gˆ k      
 N
0 for k 
 2
 cont
k  N / 2 
3) Then: dIf
 x    gˆ k e ikx

dx k  N / 2  1

4) Summary:
a) fft transform data to compute coefficients
b) scale Fourier coefficients
c) inverse fft (ifft) scaled coefficients
 Final Twist
• Matlab stores the coefficients from the fast
Fourier transform in a slightly odd order:
fˆ0 , fˆ1 ,.., fˆN , fˆN , fˆ N , fˆ N ,..., fˆ1
1  1  2
2 2 2 2

• The derivative matrix will now be a matrix

with diagonal entries:
 N N N 
i 2  0,1,..,  1, 0,   1,   2,..., 1 
 2 2 2 
 Spectral
Differentiation
Code
Typo:
See corrected
Code on webpage

1) DFT data
2) Scale Fourier
coefficients
3) IFT scaled coefficients
 Two-Dimensional Fourier
Transform
• We can now construct a Fourier expansion
in two variables for a function of two
variables.. k N / 2 jN / 2
If  x, y   fˆ eij 2 x eik 2 y  
k  N / 2  1 j  N / 2  1
jk

• The 2D inverse discrete transform and

discrete transform are:
k N / 2 jN / 2
f nm   
k  N / 2  1 j  N / 2  1
fˆjk eij 2 xn eik 2 ym for n, m  1,..., N

n  N m N
ˆf  1 N N
jk
N2
 f
n 1 m1
nm e  ij 2 xn  ik 2 ym
e for j,k      1,...,
2 2
 Advection Equation
• Recall the 2D advection equation:
C x C y C
a a 0
t x y

• We will use a Runge-Kutta time integrator

and spectral representation in space.
 Runge-Kutta Time Integrator
• We will now discuss a particularly simple
Runge-Kutta time integrator introduced by
Jameson-Schmidt-Turkel

• The idea is each time step is divided into s

substeps, which taken together
approximate the update to s’th order.
 Side note: Jameson-Schmidt-
Turkel Runge-Kutta Integrator
• Taylor’s theorem tell’s us that
 dt  d  dt 2  d 2 dt  d  
s s

C  t  dt   1      ...     C  t 
 1!  dt  2!  dt  s !  dt  

dt s1  d s1C  *
  s 1      t , t  dt 
*
t for some t
 s  1 !  dt 
• We will compute an approximate update
as:  dt  d  dt 2  d 2 dt  d  
s s

C  t  dt   1        ...     C  t 
 1!  dt  2!  dt  s !  dt  

 JST Runge-Kutta
• The numerical scheme will look like
 dt  d  dt 2  d 2 dt  d  
s s

C  t  dt   1      ...     C  t 
 1!  dt  2!  dt  s !  dt  

n 1
 2
dt  d  dt  d 
2
dt  d   n
s s

 C  1      ...     C
 1!  dt  2!  dt  s !  dt  

• We then factorize the polynomial
derivative term:
 Factorized Scheme
 dt  d  dt 2
d 
2
dts
d 
s
 n
C n1   1        ...     C
 1!  dt  2!  dt  s!  dt 
 

 n 
dt d  dt d  dt d  dt dC 
C C  C  C   .. C    
n+1 n n n n

s dt  s  1 dt  s  2 dt   1 dt   


Set C=C n
for k  s : 1:1
dt dC
CC  n

k dt
end
Cn+1 =C
 JST + Advection Equation
• We now use the PDE definition
Set C=C n
for k  s : 1:1 C C C
 ax  ay 0
C  Cn 
dt dC + t x y
k dt
end
Cn+1 =C

Set C=C n
for k  s : 1:1
dt  x C y C 
C  C  an
a
k  x y 
end
Cn+1 =C
 Now Use Spectral Representation
Set c  c
time step now consists of s substages.
for k  s : 1:1
ach stage involves:
cˆ = fft  c 
Fourier transforming the ctilde
using a fast Fourier transform (fft) dˆ x  D x cˆ
Scaling the coefficients to
differentiate in Fourier space dˆ y  D y cˆ

Transforming the derivatives back to  

d x  ifft dˆ x
physical values at the nodes by
inverse fast Fourier transform (ifft). dy  ifft  dˆ  y

dt x x
c j  c j   a j d j  a yj d yj  for j=1,..,N
Finally updating ctilde according to
the advection equation.
k
t the end we update the concentration. end
c=c
 Matlab
Implementation
• First set up the
nodes and the
differentiation
scalings.
24) Compute dt
26) Initiate
concentration
28-29) compute
advection vector
32-45) full Runge-
Kutta time step
35) fft ctilde
37) x- and y-
differentiation in
Fourier space
40-41) transform back
to physical values
of derivatives
43) Update ctilde
C x C y C
a a 0 a x   sin  2 x  cos  2 x 
t x y
a y  cos  2 x  sin  2 x 
C  t  0, x, y   e

200  x .75   y .6 
2 2

 FFT Resources
• Check out: http://www.fftw.org/ for the
fastest Fourier transform in the West

• Great list of links:

– http://www.fftw.org/links.html

• FFT for irregularly spaced data:

– http://www.math.mu-luebeck.de/potts/nfft/
 Lab Exercise
1) Download example DFT code from website.
2) Benchmark codes as a function of N
3) Write your own version using an efficient FFT
(say fftw) which:
1) FFT’s data
2) Scales coefficients
3) IFFT’s scaled coeffients
4) Write an advection code using the JST Runge-
Kutta time integrator.
This is the scalar version of Project 4
 Upshot: 2D Advection/DFT
Using MPI_Alltoall

Note: using fft the compute time will drop by approx.

A factor of 30 (for N=256)
 Upshot: 2D Advection/DFT
Using Non-Blocked Sends

Note: I isend and irecv, then compute the x-derivatives, then waitall, then compute
the y-derivatives, then isend/recv and waitall (not the best option perhaps).
 With FFT
• Replacing the DFT calls with FFT calls will
radically reduce the computation time.
• It will be much harder to hide the
communication time behind the compute
time.

• Good luck 