An Energy and Performance Efficient DVFS Scheme for Irregular Parallel Divide-and-Conquer Algorithms on the Intel SCC

The divide-and-conquer paradigm can be used to express many computationally significant problems, but an important subset of these applications is inherently load-imbalanced. Load balancing is a challenge for irregular parallel divide-and-conquer algorithms and efficiently solving these applications...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE computer architecture letters Ročník 13; číslo 1; s. 13 - 16
Hlavní autoři: Yu-Liang Chou, Shaoshan Liu, Eui-Young Chung, Gaudiot, Jeen-Luc
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York IEEE 01.01.2014
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Témata:
Algorithms
D Software/Software Engineering
D.4 Operating Systems
D.4.7 Organization and Design
D.4.7.b Distributed systems
D.4.7.f Parallel systems
D.4.8.a Measurements < D.4.8 Performance < D.4 Operating Systems < D Software/Software Engineering
Distributed prcessing
Divide & conquer algorithms
Operating systems
Parallel processing
Performance evaluation
Software engiineering
Load Imbalance
Divide-and-conquer
Intel SCC
DVFS
ISSN:1556-6056, 1556-6064
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:The divide-and-conquer paradigm can be used to express many computationally significant problems, but an important subset of these applications is inherently load-imbalanced. Load balancing is a challenge for irregular parallel divide-and-conquer algorithms and efficiently solving these applications will be a key requirement for future many-core systems. To address the load imbalance issue, instead of attempting to dynamically balancing the workloads, this paper proposes an energy and performance efficient Dynamic Voltage and Frequency Scaling (DVFS) scheduling scheme, which takes into account the load imbalance behavior exhibited by these applications. More specifically, we examine the core of the divide-and-conquer paradigm and determine that the base-case-reached point where recursion stops is a suitable place in a divide-and-conquer paradigm to apply the proposed DVFS scheme. To evaluate the proposed scheme, we implement four representative irregular parallel divide-and-conquer algorithms, tree traversal, quicksort, finding primes, and n-queens puzzle, on the Intel Single-chip Cloud Computer (SCC) many-core machine. We demonstrate that, on average, the proposed scheme can improve performance by 41% while reducing energy consumption by 36% compared to the baseline running the whole computation with the default frequency configuration (400MHz).
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1556-6056
1556-6064
DOI:10.1109/L-CA.2013.1