Prototypování paralelních aplikací založené na principi rozděl a panuj

Prototyping Parallel Applications Based on Divide and Conquer Strategy

article in a collection out of WoS and Scopus

en

The paper addresses a problem of parallel implementation of divide-and-conquer algorithms, whose performance is not always satisfactory. Questions like depth of recursion, multiple processes per processor, communication architecture and overhead, etc., are analyzed and the template for prototyping D&C algorithms in Transim language is explained. Prototypes of parallel D&C programs can be executed and their performance estimated before developing the code in detail. The technique is useful for programs with message passing as well as with shared variables. As an example a prototype of a parallel 1D-FFT benchmark targeted both to a distributed memory machine and to a SMP have been developed and simulated. The experiments show that the template can be used for quick estimation of performance and suitability of D&C approach in any given application.

Článek se zabývá problémem paralelní implementace algoritmu rozděl a panuj, jehož výkonnost není vžy uspokojující. Otázky jako hloubka rekurze, počet procesů na procesor, komunikační architektura a režie atd. jsou analyzovány a šablona v jazyce Transim pro protypování D&C algoritmu je popsána. Prototypy paralelních D&C algoritmů mohou být spuštěny a jejich výkonnost odhadnuta před implemetací. Tato technika je užitečná jak pro programy se zasíláním zpráv tak se sdílenou pamětí, což bylo demonstrováno na příkladě paralelní 1D-FFT. Experimenty ukazály, že šablona je vhodná pro rychlé odhad výkonnosti D&C řešení pro danou aplikaci.

The paper addresses a problem of parallel implementation of divide-and-conquer algorithms, whose performance is not always satisfactory. Questions like depth of recursion, multiple processes per processor, communication architecture and overhead, etc., are analyzed and the template for prototyping D&C algorithms in Transim language is explained. Prototypes of parallel D&C programs can be executed and their performance estimated before developing the code in detail. The technique is useful for programs with message passing as well as with shared variables. As an example a prototype of a parallel 1D-FFT benchmark targeted both to a distributed memory machine and to a SMP have been developed and simulated. The experiments show that the template can be used for quick estimation of performance and suitability of D&C approach in any given application.

algoritmus rozděl a panuj, paralelní výkonnost, paralelní rekurzivní FFT

Divide and conquer algorithms, parallel performance, parallel recursive FFT

2002

25.04.2002

Marq software s.r.o.

Ostrava

80-85988-71-2

Proceedings of 36th International Conference MOSIS '02 Modelling and Simulation of Systems

313–320

8