Staccato: Cache-aware work-stealing task scheduler for shared-memory systems

Результат исследований: Научные публикации в периодических изданияхстатья

2 Цитирования (Scopus)

Аннотация

Parallel tasks work-stealing schedulers yield near-optimal tasks distribution (i.e. all CPU cores are loaded equally) and have low time, memory and inter-thread synchronizations. The key idea of work-stealing strategy is that when scheduler worker runs out of tasks for execution, it start stealing tasks from the queues of other workers. It’s been shown that double ended queues based on circular arrays are effective in this scenario. They are designed with an assumption that tasks pointer are stored in these data structures, while tasks object reside in heap memory. By modifying tasks queues so that they can hold task objects instead pointers we managed to increase the performance above 2.5 times on CPU bound applications and decrease last-level cache misses 30% compared to Intel TBB and Intel/MIT Cilk work-stealing schedulers.

Язык оригиналаанглийский
Страницы (с-по)91-102
Число страниц12
ЖурналLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Том10963
DOI
СостояниеОпубликовано - 4 июл 2018
Событие18th International Conference on Computational Science and Its Applications, ICCSA 2018 - Melbourne, Австралия
Продолжительность: 2 июл 20185 июл 2018

    Fingerprint

Предметные области Scopus

  • Теоретические компьютерные науки
  • Компьютерные науки (все)

Цитировать