Strategies for Parallel Data

These explanations can be opaque—providing a means of classifying variation without any explanation of why it works— or transparent—describing what creates the observed variation. For two reasons, data mining could be the killer application that parallel computing has been seeking. First, analyzing variation appears to be algorithmically complex and hence might require levels of computing power that only parallel computers can provide in a timely way. (See the sidebar, “Why parallelize data mining?”) Second, the data sets involved are large and rapidly growing larger, and parallel computers are organized to handle such large volumes effectively, although some data-mining problems are already taxing their limits. Data-mining algorithms can potentially be parallelized in many different ways. Because designing and implementing parallel programs is expensive, it is impractical to test all of these by building implementations and comparing them. Fortunately, practical complexity measures for parallel programming are rapidly maturing. This article shows how to use such measures to assess different parallelization strategies for data-mining algorithms. We also show how the structure of some algorithms results in a doublespeedup phenomenon. A high-level cost analysis greatly simplifies the search for an effective algorithm.