The hyperdyadic index and generalized indexing and query with PIQUE

Abstract

Many scientists rely on indexing and query to identify trends and anomalies within extreme-scale scientific data. Compressed bitmap indexing (e.g., FastBit) is the go-to indexing method for many scientific datasets and query workloads. Recently, the ALACRITY compressed inverted index was shown as a viable alternative approach. Notably, though FastBit and ALACRITY employ very different data structures (inverted list vs. bitmap) and binning methods (bit-wise vs. decimal-precision), close examination reveals marked similarities in index structure. Motivated by this observation, we ask two questions. First, "Can we generalize FastBit and ALACRITY to an index model encompassing both?" And second, if so, "Can such a generalized framework enable other, new indexing methods?" This paper answers both questions in the affrmative. First, we present PIQUE, a Parallel Indexing and Query Unified Engine, based on formal mathematical decomposition of the indexing process. PIQUE factors out commonalities in indexing, employing algorithmic/data structure "plugins" to mix orthogonal indexing concepts such as FastBit compressed bitmaps with ALACRITY binning, all within one framework. Second, we define the hyperdyadic tree index, distinct from both bitmap and inverted indexes, demonstrating good index compression while maintaining high query performance. We implement the hyperdyadic tree index within PIQUE, reinforcing our unified indexing model. We conduct a performance study of the hyperdyadic tree index vs. WAH compressed bitmaps, both within PIQUE and compared to FastBit, a state-of-the-art bitmap index system. The hyperdyadic tree index shows a 1.14-1.90x storage reduction vs. compressed bitmaps, with comparable or better query performance under most scenarios tested.