OpenMP

This repo contains examples for running OpenMP programs on the ACCRE cluster.

Building Programs

Each example has a Makefile that assumes icc is used rather than gcc.

setpkgs -a intel_cluster_studio_compiler
make

Running Programs

All examples can be run with something like:

bash run.sh 2

Where 2 is the number of threads to use for parallel portions of the program.

Exercises

1. hello

• Run the program three times in a row with 8 threads. What do you notice?
• Try running the program without invoking the shell script (i.e. type ./hello). What is the default number of threads? Where does this value come from?
• Add printf("All done!\n"); at the end of the program, outside of the parallel region. What do you notice when you run the program repeatedly?

2. parallel-for

• Are variables declared inside of a parallel region defined as private or shared?
• Run your code with eight threads. How is the array split up across the threads (see hint below)? Is this behavior consistent for multiple instances of the program?
• Try changing the thread scheduling from dynamic to static - how is the array split up across threads now? Now try with the guided option (you will need to increase the loop size by an order of magnitude to see this difference clearly).

Something like:

bash run.sh 8 > out
grep -c "thread id: 0" out
grep -c "thread id: 1" out
...

might be useful for counting how data gets split up across threads.

• Run the time command on the program with a different number of threads (1, 2, 4, 8). How does the wall (i.e. real) time change?

  time bash run.sh 1 > tmp

3. parallel-for-combined-directive

• In C programs you can wrap long lines of code across multiple lines by including a backslash character at the end of a line (without a semicolon). Try splitting the long #pragma line across three lines.
• Explore some of the other clauses that parallel for loops support (https://computing.llnl.gov/tutorials/openMP/#DO). What's the difference between private, firstprivate, and lastprivate? What about the collapse clause?
• What are rules of thumb for declaring a variable as shared or private?

4. sections

• Run this code with 8 threads. How many threads end up processing data?
• Notice the absence of curly brackets {} inside each #pragma omp section block of code. Curly brackets are not needed if you are only enclosing a single line of code or a loop.
• Let's pretend we want to store the sum of arrays c and d into another array e. Where should we put that code? Are any other changes to the exisiting program needed?

5. critical

• Try commenting out the critical directive line, re-building, and running with a bunch of threads (a few hundred).

6. reduce

• Run with 2-8 threads.
• What happens if you change the operator to a minus sign in the reduction clause?

7. threadprivate

• Run with 2-8 threads.
• Note that a and x are declared threadprivate whereas b is not. How does this difference manifest itself in the output from the program?
• Rather than setting a equal to tid in the first parallel region, try commenting that line out and then initializing a to a value of 2 before the first parallel region. What output do you observe for a now? Try adding the copyin(a) clause to the end of the first parallel region. Now what's the output for a?