Question

I am planning to write a code where I want to share the work using tasks which can be split. In a serial version, I use a stack initialised with the root task. The stack is emptied by repeatedly popping a task, which is then either performed or split and the sub-tasked pushed back on the stack. What is the best (most efficient and effortless) way and interface (tbb, openmp, etc) to implement this in parallel? Is a parallel stack like this explicitly supported by any thread-parallel interface (or is there a better alternative to a stack)?

Answer 1

What you are actually describing is a work-stealing queue. This has started with the Cilk programming language (http://en.wikipedia.org/wiki/Cilk) and continues with Cilk++ that is part of the Intel Compiler. You can get a free evaluation version for Linux.

Intel TBB also features a work-stealing queue and supports automatic splitting of tasks, plus additional control if you want to split them yourself. OpenMP does not have this kind of control but I don’t think it is impossible to do.

There is also qthreads from Sandia (https://code.google.com/p/qthreads/) which a nice package that allows you to have an M:N scheduler (M user level threads on N OS threads).