Multiple Parallel Sites
You may be able to introduce parallelism
independently in more than one place in a program.
For example, consider a C/C++ program with the
general structure:
initialize(data); while (!done) { display_on_screen(data); update(data); }
You might be able to parallelize the display and
update operations independently:
display_on_screen(data) { ANNOTATE_SITE_BEGIN(site_display); for (each block of data) { ANNOTATE_ITERATION_TASK(task_display); display the block of data; } ANNOTATE_SITE_END(); } update(data) { ANNOTATE_SITE_BEGIN(site_update); for (each block of data) { ANNOTATE_ITERATION_TASK(task_update); update the block of data; } ANNOTATE_SITE_END(); }
Each iteration of the main loop would still do the
display and then the update, but the display and update operations could be
performed much faster.
Depending on your program, you need to decide whether
to implement multiple parallel sites at the same or at different times:
- When two parallel sites are truly disjoint or have overlapping functions that are purely functional and do not show problems reported by the Dependencies tool, you can consider parallelizing those sites separately at different times.
- When considering multiple parallel sites that overlap on the same call trees - such as multiple sites that call the same (common) utility functions - consider parallelizing or not parallelizing the entire set of parallel sites at the same time.
You need to determine the cause of each dependency
and fix it. If you have multiple parallel sites that overlap on the same call
trees - such as multiple sites that call the same utility functions (common
code) - read the help topic Fixing Problems in Code Used by Multiple Parallel
Sites.