MATLAB Parallel Computing Toolbox (PCT)
How to use pmode
pmode is an interactive parallel environment enhanced with a GUI. For a small number of workers
, each has its own dedicated sub-display-window. This is an excellent tool
for learning the PCT and for parallel code development and debugging. Starting pmode
automatically launches an interactive batch job that allocates the requested workers.
-
To start pmode (with default configuration)
>> pmode start % "help pmode" for more details
Starting pmode using the parallel configuration 'SGE'.
Job output will be written to: /usr6/kadin/examples/pct/Job3.mpiexec.out
QSUB output: Your job 58551 ("Job3") has been submitted
Waiting for parallel job to start...
Connected to a pmode session with 4 labs.
In the above, pmode was invoked with default configuration, which is
defined as the Katana cluster Sun Grid Engine (SGE) batch scheduler. The last printed
line indicated that 4 labs (workers) have been assigned. You could override the default
with ">> pmode start 8" to request 8 labs.
A separate GUI window will pop up which consists of a group of sub-windows
displaying data
on the local labs in response to commands issued at P>> at the bottom of
this window. The P>> prompt distinguishes itself from the regular
MATLAB prompt (>>) as being dedicated to pmode.
To end pmode, do either of the below:
>> pmode quit % quit|exit|close from the MATLAB prompt
P>> exit % exit the Parallel Command Window
-
To copy data from a worker (lab) to the client (MATLAB workspace)
P>> pmode lab2client y 3 yc % copy y of lab 3 to yc on client
-
To copy data from client to workers
P>> pmode client2lab yc 2 y % copy yc on client to y on lab 2
-
Make plots
pmode environment does not support graphics display. To make
plots:
- Copy data from the workers to the MATLAB client
- Make plots in the MATLAB window (>>).
Here is an example.
P>> x = 2*labindex; % [2 4 6 8]
P>> D = zeros(1,numlabs,codistributor()); % create a shared array
P>> for i=drange(1:numlabs), D(i) = x; end % saves x in D according to labindex
P>> y = gather(D, 1); % gather from all labs to lab 1
P>> pmode lab2client y 1 yc % pmode lab2client labvar lab clientvar
Now, switch to the MATLAB window.
>> yc % verify the content of yc in the MATLAB window
yc = 2 4 6 8
>> plot(yc) % plot a line in the MATLAB window
Alternatively, D can be generated without using drange:
P>> D = ones(1,numlabs,codistributor()) * x;
- matlabpool + spmd is equivalent to pmode but without
a graphics interface or a few other commands such as whos.
- If you are developing a procedure via pmode, once all the steps have been
worked out, you can capture the history by selecting lines on the command list on the left pane of the
pmode window and save it into a script file for batch applications. Specifically, to select consecutive
lines, press the shift key and left-mouse click the beginning and
end lines. To select arbitrary lines, press the control key and
use the left-mouse button to select the individual lines. It is important to note
that pmode operations and matlabpool + spmd operations are not
completely conformal. Alternatively, diary can be used to
capture the commands exercised. Some conversion may be necessary to turn
pmode-generated commands into matlabpool + spmd commands.
-
Example -- Solving a Linear Algebraic System of Equations
n = 3; % define matrix size
M = rand(n,n,codistributor()); % create distributed random number matrix
A = M+M'; % A = M*M' is a real, symmetric matrix
x = [1:n]'; % the solution column vector x
x = codistributed(x,'convert'); % distribute x among labs
b = A*x; % b = Ax is the right-hand-side
x2 = A\b; % solves for x2
whos % whos shows that all arrays are distributed
[V, D] = eig(A); % find eigen values of real, symmetric matrix
|
