OpenMP Tutorial

This directory contains files to support a short, three-day overview of OpenMP, a standard for multi-threaded, parallel computing in C++. The tutorial is organized by exercise.

Exercise 1 - Hello, World

Use OpenMP to produce a program that displays several messages of the form Hello, from thread 1 of 4. This will require that students use the parallel directive, access the thread ID, and access the number of threads. To be interesting, students should also modify the number of threads.

Introduces

• concept of fork/join model
• parallel directive
• <omp.h>
• omp_get_num_threads
• omp_get_thread_num
• OMP_NUM_THREADS

Exercise 2 - c = a + b

Given two std::vectors a and b, compute their element-wise sum c in parallel using just the parallel directive, the thread ID, and the number of threads.

Introduces

• omp_get_wtime
• concept of "work sharing"
• concept of shared memory

Exercise 3 - c = a + b (the easy way)

Repeat Exercise 2 by using the for directive, thereby eliminating any need for explicit work sharing.

Introduces

• for directive

Exercise 4 - The Parallel Dot Product

Use the parallel and for directives to compute the dot product of two arrays a and b. Assess the parallel efficiency for various thread counts and various problem sizes.

Introduces

• concept of private memory
• concept of race conditions
• critical directive
• atomic directive
• reduce directive

Exercise 5 - The Parallel Sum of 1, 2, ..., n

Use the same basic structure from Exercise 4 to compute the sum of 1, 2, ..., n. How does the performance differ from those of Exercise 4?

Introduces

• concept of memory bandwidth limitations

Exercise 6 - Heat Conduction in 2-D with Jacobi Iteration

You are given a basic program that computes the temperature in a two-dimensional box subject to fixed boundary temperatures. Your job is to parallelize this code and to extract the best performance possible.