Parallel Weighted Random Sampling

Data structures for efficient sampling from a set of weighted items are an important building block many applications. However, few parallel solutions known. We close these gaps. give efficient, fast, and practicable distributed algorithms data that support single (alias tables, compressed structures). This also yields simplified more space-efficient sequential algorithm alias table construction. Our approaches to k out n with/without replacement subset (Poisson) output-sensitive , i.e., the use work linear in number different samples. is interesting case. Weighted random permutation can be done by sorting appropriate deviates. show this possible with work. Finally, we communication-efficient, highly scalable approach (weighted unweighted) reservoir sampling. based on fully model streaming might independent interest. Experiments tables near speedups using up 158 threads shared-memory machines. An experimental evaluation 5,120 cores shows good speedups.