Core Concepts
Predicting parallel performance using runtime distributions.
Abstract
The paper analyzes the scalability and parallelization of local search algorithms for the Satisfiability problem.
A framework is proposed to estimate parallel performance by analyzing the runtime behavior of sequential versions.
Empirical data on Sparrow and CCASAT solvers up to 384 cores is compared to predicted performances.
Different types of instances exhibit varying behaviors with different runtime distributions.
The study shows that the model accurately predicts parallel performance close to empirical data.
Experimental results indicate that Sparrow scales better than CCASAT in most cases, with linear speedup observed for crafted instances.
Theoretical distributions like exponential, lognormal, shifted exponential are used to characterize runtime behaviors.
Stats
ランダムなインスタンスに対するSparrowの最小ランタイムは98.0、最大ランタイムは7946.3です。
CCASATの最小ランタイムは103.6で、最大ランタイムは1177.9です。
クラフトされたインスタンスにおけるSparrowの平均ランタイムは3440.3です。
Quotes
"Most researchers focus on developing parallel portfolios for multi-core architectures."
"Empirical speedup factors for both solvers are far from linear."
"The study shows that not all combinatorial problems exhibit perfect exponential behavior."