Code Parallelization

  2. /
  3. Training
  4. /
  5. Online training
  6. /
  7. Code Parallelization

<back

image
Code Parallelization/Hybridization
Subtitle

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer lorem justo, sagittis id risus a, sodales sodales lacus. Donec ex massa, malesuada ut porta et, ornare non turpis. Donec orci dui, sodales quis felis a, auctor ullamcorper lorem. Donec posuere hendrerit maximus. Aenean placerat tempus porttitor.

Transport problem

A serial code evolving the equation of motion:     d/dx + d/dy = -d/dt, is provided.

image

Temperature values are calculated over a grid and initialized with a gaussian distribution. The points of the grid represent the local indexes (ix, iy) of the matrix that contains the temperature values. The domain is shown in orange in the picture below. Boundaries are represented in white. Data are evolved along Y=X direction, i.e. towards the up-right corner of the coordinate system.

Compile the serial code, for example with the INTEL compiler (or the GNU compiler adding the argument “-lm” to the compilation line to avoid undefined reference to “exp” calls).

image

The execution of the code produces two files: ‘transport.dat’, the data set at time t=0, and ‘transport_end.dat’, the data after the transport dynamics. Data can be visualized with the command:

(please log in with the command “ssh -X [email protected]” note the option -X)

module load gnuplot echo “set hidden3d; splot ‘transport.dat’ w l, ‘transport_end.dat’ w l” | gnuplot -persist

Exercise 1

Parallelize the code by using domain decomposition technique: let’s divide the domain in slices along the y coordinate.

Keep in mind that the data is distributed by “leading dimension”:
FORTRAN: first coordinate
C: last coordinate

The important thing is that the data to send/recv are contiguous in memory. Otherwise a copy of them to/from a temporary contiguous buffer is needed.

Optional tasks (easy for the serial version… try them in parallel too):
– Evaluate the average over all the system and check it every 10 steps
– Find the global maximum and print its position every 10 steps

HINTS:  
MPI_PROC_NULL communications to/from MPI_PROC_NULL do nothing

 

C
iy2y

this function has to deal with the domain decomposition along the y axis
iy2y(iy, <…>, <…>)

 

 

FORTRAN
iy2y

this function has to deal with the domain decomposition along the y axis
iy2y(iy, <…>)
comms

use a module to make variables like “nprocs, … ” available to the routines of the program
    Solutions

Exercise 2

– Add the OpenMP directives to the MPI code to parallelize some loops and to manage the MPI communications.
– Select and check the right MPI level of thread support.
– Print both the process and thread identifiers.
– Compile your code with the OpenMP support.
– Run with different configurations for processes and threads.

    Solutions

Exercise 3 (optional)

Try to hybridize the solution of exercise 15 of the MPI tutorial, about the matrix transposal (http://www.www.hpc.cineca.it/content/solution-15), by adding OpenMP directives. Remember to use “omp master” directive when a task should be performed by only one thread-per-task.

    Solutions
MENU