Hi, I was wondering how this could be done:We have different parts of the software that have different priorities. For instance lets imagine we have a user thread and a low-level processing thread. Each of them may call parallel algorithms (sort, for, for_each, pipelines, ...) We would like to have user-level threads have more priority when executing parallel tasks.What I was thinking is something like that: set a priority on a "master" thread and dependant sub-tasks gets priority queued accordingly within tbb. Therefore, worker threads would pop items from the task queue in order or priority.Or maybe it is the other way around and it is masters that asks for tasks to perform and dispatch.. Therefore, a master with a higher priority would have more chances to get CPU time and therefore send job to workers threads more often..Mike
assigning a priority to a task/algorithm
For more complete information about compiler optimizations, see our Optimization Notice.
- Log in to post comments
Fri, 02/03/2012 - 15:23
TBB is not configured for multiple threadpriorities, although there is some limited control for task priority.
You can have non-TBB threads incorporated into your program that run at different priority levels. But be careful as to which domain your code is running in. Cross domain calls my yield unexpected results.
Jim Dempsey
- Log in to post comments
Sun, 02/05/2012 - 00:27
Hi Mike,there is a blog about tbb task priorities. I guess this is what are you asking abouthttp://software.intel.com/en-us/blogs/2011/04/01/task-and-task-group-priorities-in-tbb/--Vladimir
- Log in to post comments
Sat, 03/10/2012 - 14:01
Quoting mlemay@coveo.comFYI, parallel_sort does not have prototypes with task_group_context parameter.. (I'm using 4.0U2, 278oss)
Thanks for bringing this to attention.
Initially, task_group_context parameter was used to cancel some algorithm from outside, and we thought that there is no practical sense to cancel sort, which would result in the array being modified (in place)but still unsorted.
However, now with the context extended to also specify priority, adding a corresponding overload to parallel_sort certainly is desirable.
