Autotuning and Self-Adaptability in Concurrency Libraries

TL;DR

This paper presents an extension to the TBB library that enables automatic autotuning of parallel applications with minimal code modifications, significantly improving performance in some cases.

Contribution

The authors introduce an autotuning extension to TBB that automates performance optimization without requiring extensive code changes.

Findings

Some applications see up to 28% speedup with autotuning.

Autotuning reduces manual effort in performance optimization.

Performance gains vary across different applications.

Abstract

Autotuning is an established technique for optimizing the performance of parallel applications. However, programmers must prepare applications for autotuning, which is tedious and error prone coding work. We demonstrate how applications become ready for autotuning with few or no modifications by extending Threading Building Blocks (TBB), a library for parallel programming, with autotuning. The extended TBB library optimizes all application-independent tuning parameters fully automatically. We compare manual effort, autotuning overhead and performance gains on 17 examples. While some examples benefit only slightly, others speed up by 28% over standard TBB.