The granularity control challenge
One central difficulty in parallelism is to spawn sequential subtasks of the right size. On the one hand, creating sequential tasks that are too small leads to unacceptable overheads. On the other hand, creating sequential tasks that are too big caps the number of cores that can be used.
The traditional approach to granularity control consists in deciding whether to sequentialize subtasks or not based on a cutoff value hard-coded in the source code, however this approach is not portable at all. Another approach is auto-tuning, which consists in automatically trying various possible values for the cutoff on a given hardware, but this takes time and requires samples of input data. The goal is thus to come up with a portable, online approach to granularity control.
We have developed a new approach to granularity control that combines asymptotic complexity annotations with runtime profiling. We require the programmer to annotate his parallel functions with an asymptotic complexity expression. We then use runtime profiling for deducing the constant factors that apply. Using this information, we are able to predict execution time and enforce our scheduling policy: any subtask that is predicted to take less than a fixed amount of time gets sequentialized. This approach works for any divide-and-conquer algorithm whose worst-case complexity matches its average complexity.
We have proved bounds showing that our granularity control strategy leads to provably-good parallel run times. Moreover, we have implemented our approach and shown that it works well in practice.
Umut A. Acar, Arthur Charguéraud, and Mike Rainey
OOPSLA, October 2011
We first implemented these ideas in the Manticore compiler (SML syntax), as a source-to-source translation. We later re-implemented them in our C++ library PASL, where asymptotic cost annotations are to be provided in the form of member functions of the closure objects.