Smart access strategies for Data-­Centric processing

Abstract

Data movement in analytical database systems is a critical bottleneck, driving energy consumption and infrastructure costs. In the context of storage access, this thesis contributes techniques to mitigate these costs through Cooperative Refinement, the symbiotic interplay between indexing and data-centric processing.

First, we address the intersection of these two fields: We by analyze probabilistic data structures, such as Bloom filters and binary sketches, as candidates for Processing-in-NAND (PiN) due to their high error tolerance. To expand the applicability of current PiN architectures, we propose a scheme for emulating inequality comparisons inside NAND. To maximize the potential of fine-granular index information, we present early work on Gravity Store, a data-centric in-storage materialization engine for declarative analytics.

The second major contribution is Team-based indexing, a generalization of bitmap indexing for selective, high-dimensional range queries. By forming "Teams" of moderately-sized attribute subsets, this strategy improves runtime efficiency and reduces storage overhead compared to traditional indexing. We address the central challenges of efficient index intersection and Team composition. Finally, we introduce TeamBench, a benchmark generator specifically designed to evaluate these index intersection performances at scale.