CppCon 2015

If you've used a C++ parallel-programming system in the last decade, you've probably run across the term "work stealing." Work stealing is a scheduling strategy that automatically balances a parallel workload among available CPUs in a multi-core computer, using computation resources with theoretical utilization that is nearly optimal. Modern C++ parallel template libraries such as Intel(R)'s TBB or Microsoft*'s PPL and language extensions such as Intel(R) Cilk(tm) Plus or OpenMP tasks are implemented using work-stealing runtime libraries.

Most C++ programmers pride themselves on understanding how their programs execute on the underlying machine. Yet, when it comes to parallel programming, many programmers mistakenly believe that if you understand threads, then you understand parallel runtime libraries. In this talk, we'll investigate how work-stealing applies to the semantics of a parallel C++ program. We'll look at the theoretical underpinnings of work-stealing, now it achieves near optimal machine utilization, and a bit about how it's implemented. In the process, we'll discover some pit-falls and how to avoid them. You should leave this talk with a deeper appreciation of how parallel software runs on real systems.

Previous experience with parallel programming is helpful but not required. A medium level of expertise in C++ is assumed.