随机算法异步并行化的效率分析

Efficiency Analysis of Asynchronic Parallelization of Randomized Algorithms

随机算法的执行时间具有不确定性,这种不确定性为随机算法的异步并行提供了良好的基础,已有许多计算实验表明了随机算法的异步并行可以达到线性甚至超线性的加速.对于求解SAT问题的随机算法RDP,研究了异步并行效率与运行时间分布和处理器数目之间的关系.应用一种单峰分布──分段线性分布模型来模拟随机算法的运行时间分布.理论分析和计算结果均表明:当处理器数目k较小和单峰位于分布的前部时,随机算法的异步并行具有近线性加速.

The uncertainty of running time of randomized algorithms provides a better opportunity for asynchronized parallelization. There are many computing experiments verifying that the asynchronized parallelizating acceleration of randomized algorithms are linear or even superlinear. For randomized algorithm RDP solving for SAT (satisfiability) problem, the relation among efficiency of asynchronized parallelization, distribution of running time and number of processors are investigated. A model of piecewise-linear distribution is applied to simulate the running time distribution of randomized algorithms. This model of distribution is a kind of single peak. Both theoretical analysis and computing experiment indicates that asynchronized parallelization of randomized algorithms are of near linear acceleration when the processors are less and the single peak is located near the front of running time distributions.