Parallel K-Tree: A multicore, multinode solution to extreme clustering

Clustering is a popular technique that can help make large datasets more manageable and usable by grouping together similar objects. Most clustering approaches are too computationally expensive for datasets that are very large or complex. This work presents Parallel K-Tree, a hierarchical data structure and clustering algorithm that takes advantage of modern computing environments to cluster extremely large datasets. Parallel K-Tree produces high-quality clusters and scales more efficiently than traditional, parallelized and state-of-the-art approaches. Parallel K-Tree was applied to a large (8 terabyte) collection of Landsat 5 satellite images. This required clustering of 540 billion objects into eight billion clusters — a two orders of magnitude size increase over any reported alternative approach. Furthermore, Parallel K-Tree was executed on just two commodity servers — rather than a high-performance supercomputer.