Experience from runtime implementations: scheduling and parallelism


Tullio Vardanega

Presentation title

Experience from runtime implementations: scheduling and parallelism

Authors

Tullio Vardanega

Institution(s)

University of Padua

Presentation type

Technical presentation

Abstract

The proposed presentation has two parts, which reflect recent work by members of the real-time research group I lead at UNIPD.

Part I: FPS vs EDF: the judgment day reversed

Where the seminal work by Liu and Layland in 1973 sustained the superiority of EDF to FPS (RM) for single-processor systems in point of theory, results published in 2005 compounded that argument with experimental evidence. Having access to a lean and efficient FPS with IPCP locking runtime, from an implementation of the Ada Ravenscar Profile for the Leon processor, we developed a variant of it that implements EDF with DFP locking, which could be transparently attached to the same set of Ravenscar applications. With those tools in hand, we performed an extensive quantitative comparison between those two runtimes and obtained interesting results, which kind of reverse the so-called judgment day.

Part II: Seeking time-composable parallelism

Modern many-core platforms provide massive computational capacity that in principle enable the user to unleash numerous opportunities for extensive parallelization.

This trend not only challenges the way the application software is designed (for it must be able to take benefit of those opportunities, which is far from trivial), but also the way the runtime abstracts the complex hardware fabric underneath (for it must neither overwhelm the application with obscure options nor stand in the way of the application's intent).

We focus on the Kalray MPPA, a recent manycore design developed for scalable compute-intensive power-efficient embedded applications, to demonstrate how opportune combination and adaption of state-of-the-art solutions, implemented in a bare-metal parallelization library that we developed from scratch, allow improving the efficiency and the predictability of the system significantly.


Additional material

  • Presentation slides: [pdf]

  • Warning: Undefined variable $ADDITIONAL_MATERIAL in /var/www/html/iwes/2017/presentations.phtml on line 79