Data-driven taskification instead of trial & error

The introduction of tasks (taskification) is, in many projects, a trial and error process that often fails to deliver the performance that developers would expect or hope for. Reasons are too tiny tasks, the wrong choice of tasks vs data-parallel or serial code parts, or task runtimes which do not perform. We propose a data-driven approach where developers first annotate `what could be a task' and then create a taskification fingerprint by running their code. This fingerprint is then piped through a task simulator which delivers taskification recommendations and performance predictions. The latter can be compared to obtained performance and uncover task runtime flaws, while the overall process becomes an iterative, yet well-defined engineering approach.

---

Contributors:

• Tobias Weinzierl (Durham University)
• Marion Weinzierl (Durham University)
• Aidan Chalk (Hartree Centre, STFC Daresbury Laboratory)
• Rupert Ford (Hartree Centre, STFC Daresbury Laboratory)
• Adam Tuft (Durham University)