Parallelization of Bin Packing on Multicore Systems

We study effective parallelization of approximation algorithms for the one-dimensional bin packing problem on a multicore platform. Bin packing is a classic combinatorial optimization problem that aims to pack a given sequence of items into a minimum number of equal-sized bins. The problem potential...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2016 IEEE 23rd International Conference on High Performance Computing (HiPC) s. 311 - 320
Hlavní autoři: Ghosh, Sayan, Gebremedhin, Assefaw H.
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 01.12.2016
Témata:
Algorithm design and analysis
Approximation algorithms
Bin packing
Combinatorial optimization problems
Dynamic scheduling
First Fit Decreasing
Heuristic algorithms
Multicore processing
OpenMP
Parallel algorithms
Upper bound
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:We study effective parallelization of approximation algorithms for the one-dimensional bin packing problem on a multicore platform. Bin packing is a classic combinatorial optimization problem that aims to pack a given sequence of items into a minimum number of equal-sized bins. The problem potentially serves as a model for a wide variety of applications. Examples include: packing data into chunks in a memory hierarchy in a given system to increase application performance, loading vehicles subject to weight limitations, and packing TV commercials into station breaks. Bin packing has long served as a proving ground for the analysis of approximation algorithms and played a crucial role in the development of much of the theory of approximation algorithms. Its parallelization, however, has received comparatively much less attention. In this work, we develop multiple parallel versions of an effective approximation algorithm (First Fit Decreasing) for the problem and investigate the trade-off between solution quality and execution time. We use OpenMP and Cilk Plus as mechanisms for achieving the parallelization. The new parallel algorithms obtain a speedup of more than 10× (on 32 cores) for moderate to large input sequences without sacrificing much on the quality of solution produced by the sequential algorithm - in particular, we see only about 3 to 30% increase in the number of bins compared to the sequential version. In turn, the solution obtained by the sequential First Fit Decreasing algorithm is provably almost optimal (the approximation ratio is less than 1.3).
DOI:10.1109/HiPC.2016.044