Autotuning of a Cut-Off for Task Parallel Programs

A task parallel programming model is regarded as one of the promising parallel programming models with dynamic load balancing. Since this model supports hierarchical parallelism, it is suitable for parallel divide-and-conquer algorithms. Most naive divide-and-conquer task parallel programs, however,...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Proceedings, IEEE 10th International Symposium on Embedded Multicore/Many-core Systems-on-Chip : 21-23 September 2016, Lyon, France S. 353 - 360
Hauptverfasser: Iwasaki, Shintaro, Taura, Kenjiro
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 01.09.2016
Schlagworte:
Algorithm design and analysis
autotuning
compilers
cut-off
Heuristic algorithms
Load modeling
Optimization
Parallel processing
Parallel programming
performance optimization
task parallelism
Vegetation
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:A task parallel programming model is regarded as one of the promising parallel programming models with dynamic load balancing. Since this model supports hierarchical parallelism, it is suitable for parallel divide-and-conquer algorithms. Most naive divide-and-conquer task parallel programs, however, suffer from a high tasking overhead because they tend to create too fine-grained tasks. There are two key idea to enhance the performance of such a program: serializing a task in a cut-off condition which is a tradeoff between decrease of concurrency and parallelization overheads, and applying effective transformations for the task in the condition. Both are sensitive to algorithm features, rendering optimization solely with a compiler ineffective in some cases. To address this problem, we proposed an autotuning framework for divide-and-conquer task parallel programs. It automatically searches for the optimal combination of three basic transformation methods and switching conditions with less programmers' efforts. We implemented it as an optimization pass in LLVM. The evaluation shows the significant performance improvement (from 1.5x to 228x) over the original naive task parallel programs. Moreover, it demonstrates the absolute performance obtained by our autotuning framework was comparable to that of loop parallel programs.
DOI:10.1109/MCSoC.2016.51