Implementation of a work stealing queue in C/C++? [closed]

Answer 1

The closest implementation of this work stealing algorithm I've found is something called Wool by Karl-Filip Faxén. src / report / comparison

Answer 2

OpenMP may very well support work-stealing although its called recursive parallelism

OpenMP forum post

The OpenMP specification defines tasking constructs (which can be nested, so are very suitable for recursive parallelism) but does not specify the details of how they how they are implemented. OpenMP implementations, including gcc, typically use some form of work stealing for tasks, though the exact algorithm (and the resulting performance) may vary!

See #pragma omp task and #pragma omp taskwait

Update

Chapter 9 of the book C++ Concurrency in Action describes how to implement "work stealing for pool threads". I haven't read/implemented it myself but it doesn't look too difficult.

Answer 3

No free lunch.

Please take a look the original work stealing paper. This paper is hard to understand. I know that paper contains theoretical proof rather than pseudo code. However, there is simply no such much more simple version than TBB. If any, it won't give optimal performance. Work stealing itself incurs some amount of overhead, so optimizations and tricks are quite important. Especially, dequeues are must be thread-safe. Implementing highly scalable and low-overhead synchronizations are challenging.

I'm really wondering why you need it. I think that proper implementation means something like TBB and Cilk. Again, work stealing is hard to implement.

Answer 4

There exist a tool to simply doing it in an very elegant way. It is a really effective way to parrallelize your program in a very short time.

Cilk project

HPC Challenge Award

Our Cilk entry for the HPC Challenge Class 2 award won the 2006 award for ``Best Combination of Elegance and Performance''. The award was made at SC'06 in Tampa on November 14 2006.