Experimental evidence for the power of random sampling in practical parallel algorithms

Recent results in parallel algorithm theory have shown random sampling to be a powerful technique for achieving efficient bounds on the expected asymptotic running time of parallel algorithms for a number of important problems. The authors show experimentally that randomization is also a powerful practical technique in the design and implementation of parallel algorithms. Random sampling can be used to design parallel algorithms with fast expected run times, which meet or beat the run times of methods based on more conventional methods for a variety of benchmark tests. The constant factors of proportionality in the run times are small, and, most importantly, the expected work (and hence running time) avoids worst cases due to input distribution. They justify the approach through experimental results obtained on a Connection Machine CM-2 for a specific problem, namely, segment intersection reporting, and explore the effect of varying the parameters of the method.<>