Homework 5, due 08.00, 15/10-2014

Parallel Differentiating and integration with OpenPM

In this homework you will use your code from homework 4 to build a program that uses shared memory to differentiate and integrate. Most likely your computer has OpenMP so you should be able to do most of the development locally and just run the timing tests on Stampede.

  • First! pick one mapping from the last homework (for example a quarter annulus) and check that your code converges with second order.
  • Next, use OpenMP constructs to parallelize your serial code. Make sure you do this in a non-intrusive way so that the code still compiles in serial mode. Try to make your program as parallel as possible, I will take this into account when grading.
  • Demonstrate that your code gives the same result independent of number of threads used.
  • Use the omp_get_wtime() function to time the computational part of your code (try to exclude allocate statements but include assignments, where you can use the workshare construct).
  • Before starting the timing look over your code for places where you can optimize it by the techniques discussed in class. Also try out some different compiler options. See this and this.
  • For simplicity set the number of gridpoints the same in both directions and time your program with 1 to 16 cores for grids of size 20 by 20 to 800 by 800 with increments of 1. Display the results in a figure where you scale the wall clock time by the inverse of the size of the grid, \(n_r \times n_s\). This should give you roughly straight lines with jumps when you hit cache size limits.
  • Weak scaling. Compute the speedup and efficiency (1-16 cores) for fixed problem size. Try a small grid and a larger grid.
  • Strong scaling. Compute the speedup and efficiency (1-16 cores) for grids with a fixed number of gridpoints per core.
  • Count (by hand by looking at the code) the number of floating point operations you perform in the differentiation and add a timing to that part of the code. For the 800 by 800 case report how many floating points per second you get and compare to the theoretical maximum.

As usual, arrange your results neatly in your report and comment on them. This time, also discuss the ways you made your code parallel and what, if anything, you could improve.