Paper Info
Title
Parallelization of Divide-and-Conquer Applications on Intel Xeon Phi with an OpenMP Based Framework.
Abstract
The paper proposes an approach for efficient and flexible parallelization of divide-and-conquer computations using the modern Intel Xeon Phi accelerators. Many real-life problems follow the divide-and-conquer paradigm and consequently generate either balanced or imbalanced trees. The paper proposes an OpenMP multi-threaded implementation of a general framework that requires coding basic divide-and-conquer constructs such as data partitioning, computations and result integration. Mapping computations onto threads is handled by the underlying runtime layer. The paper presents performance results for a parallel adaptive quadrature integration resulting in an irregular and imbalanced tree. It is shown that speed-ups obtained reach around 90 for parallelization of an irregular adaptive integration code compared to maximum speed-ups of 98 for code without thread management at various levels of the divide-and-conquer tree. Results for various thread affinities are shown. The framework, for which the source code is presented, can be easily reused for any other divide-and-conquer application.
Year
DOI
Venue
2015
10.1007/978-3-319-28564-1_9
INFORMATION SYSTEMS ARCHITECTURE AND TECHNOLOGY, ISAT 2015, PT III
Keywords
DocType
Volume
Divide-and-conquer,Parallelization,Adaptive integration,Intel Xeon Phi
Conference
431
ISSN
Citations 
PageRank 
2194-5357
1
0.37
References 
Authors
0
1
Name
Order
Citations
PageRank
Pawel Czarnul112121.11