Engineering Parallel Applications with Tunable Architectures

  • Tagung:

    Vortrag 

  • Autoren:

    Walter F. Tichy

  • Summary

    Current multicore computers differ in many hardware char-acteristics. Software developers thus hand-tune their par-allel programs for a specific platform to achieve the bestperformance this is tedious and leads to non-portable code.Although the software architecture also requires adaptationto achieve best performance, it is rarely modified becauseof the additional implementation effort. The Tunable Ar-chitectures approach proposed in this paper automates thearchitecture adaptation of parallel programs and uses anauto-tuner to find the best-performing software architecturefor a particular machine. We introduce a new architecturedescription language based on parallel patterns and a frame-work to express architecture variants in a generic way. Sev-eral case studies demonstrate significant performance im-provements due to architecture tuning and show the appli-cability of our approach to industrial applications. Softwaredevelopers are exposed to less parallel programming com-plexity, thus making the approach attractive for experts aswell as inexperienced parallel programmers.
  • Jahr:

    2010 

  • Links:
Beteiligte Mitarbeiter (zufällige Reihenfolge)
Titel Vorname Nachname
Prof. em. Dr. Walter F. Tichy

Bibtex

@vortrag{,
author={Walter F. Tichy},
title={Engineering Parallel Applications with Tunable Architectures},
year=2010,
month=May,
url={https://ps.ipd.kit.edu/downloads/vor_2010_engineering_parallel_applications_tunable_architectures.pdf},
abstract={Current multicore computers differ in many hardware char-acteristics. Software developers thus hand-tune their par-allel programs for a specific platform to achieve the bestperformance this is tedious and leads to non-portable code.Although the software architecture also requires adaptationto achieve best performance, it is rarely modified becauseof the additional implementation effort. The Tunable Ar-chitectures approach proposed in this paper automates thearchitecture adaptation of parallel programs and uses anauto-tuner to find the best-performing software architecturefor a particular machine. We introduce a new architecturedescription language based on parallel patterns and a frame-work to express architecture variants in a generic way. Sev-eral case studies demonstrate significant performance im-provements due to architecture tuning and show the appli-cability of our approach to industrial applications. Softwaredevelopers are exposed to less parallel programming com-plexity, thus making the approach attractive for experts aswell as inexperienced parallel programmers.},
organization={Universität Augsburg},
}