OpenMP: Work sharing and controlling thread data

Overview

Teaching: 0 min
Exercises: 0 min

Questions

• How do OpenMP threads implicitly and explicitly share work among themsleves?

• How do data are managed?

Objectives

• Understand the basic work sharing constructs: for, secions, and single.

• Understand thread-based data sharing in OpenMP

• Understand the outcomes of OpenMP programs with different work sharing constructs.

1. Work sharing constructs

• OpenMP utilizes work sharing constructs to facilitate dividing parallelizable work among a number of threads.
• The work sharing constructs are:
  • for: divide loop iterations among threads.
  • sections: divide sections of codes among themselves.
  • single: the section is executed by a single thread.

2. Work sharing construct: sections

• Used when parallelize predetermined number of independent work units.
• Within a primary sections construct, there can be multiple section construct.
• A section can be executed by any available thread in the current team, including having multiple sections done by the same thread.

3. Hands on: sections

• In the EXPLORER window, double-click on csc466/openmp and select New Dir to create a new directory in openmp called sections.
• Inside sections, create a file named hello_sections.c with the following contents:

4. Challenge

Given the following functions: y=x⁴ + 15x³ + 10x² + 2x
develop an OpenMP program called poly_openmp.c with sections/section directives. Each section should handle the calculations for one term of the polynomial.

Solution

5. Work sharing construct: single

• Limits the execution of a block to a single thread.
• All other threads will skip the execution of this block but wait until the block is finished before moving on.
• To enable proceed without waiting, a nowait clause can be added.

6. Hands on: single

• Inside sections, create the following files:

hello_sections_nosingle.c:

hello_sections_single.c:

hello_sections_single_nowait.c:

Compile and run the above files:

7. Shared and private data

• Data declared outside of a parallel region will be shared among all threads.
• Data declared inside of a parallel region will be private to individual thread.

8. Hands on: problem with shared data

• Inside sections, create a file named counter_openmp.c with the following contents:

Key Points