Search Algorithm Portfolios with Multithreaded Components

Takumi Shimoda, Alex S. Fukunaga

Track: Short Paper (5 pages including references)

Previous Publication: No, the submission has not been published or accepted at another conference.

Keywords: Satisficing search, Parallelization, portfolio

Abstract:

One successful approach to parallel satisficing search is a search algorithm portfolio, where each processor executes an independent search process. Portfolios are competitive with parallel search algorithms where threads use shared Open/Closed lists. We investigate the scalability of portfolios vs. parallel search algorithm with shared Open/Closed, and show that the benefits due to diversification of portfolio searches hit a plateau, after which adding additional independent components yield diminishing marginal returns.
We show that hybrid portfolio/cooperative algorithms where each portfolio component is a multi-threaded, shared Open/Closed search algorithm can significantly outperform both pure portfolios as well as pure shared Open/Closed parallel search algorithms.

Submission Number: 8