In memory data management : from hardware to application

Abstract

With the availability of very large and inexpensive main memory, it is becoming practical to manage data managements in main memory and benefit from high-speed access. For instance, in-memory database management systems (e.g., SAP HANA and Oracle TimesTen) provide much higher performance over disk-oriented database management systems for relational data. In this thesis, we identify and address some unsolved issues in in-memory data management, from hardware to applications. First, we exploit the hardware aspect (e.g., CPU and memory) to accelerate distance computations (on data points), which are core subroutines in many applications, e.g., trajectory search, motif discovery and kNN classification. This involves two research problems: (1) how to exploit every CPU cycle for computation, and (2) how to exploit every bit of main memory for caching data points. Our work is orthogonal to existing pruning techniques and index structures on data points. Regarding (1), we unlock the potentials of modern commodity CPUs (i.e., data parallelism, CPU caches, branch prediction). Regarding (2), we propose to cache compact approximate representations of data points in main memory in order to reduce the candidate refinement time in existing kNN search methods. For each research problem above, we evaluate the performance of our solutions on real datasets and show that our solutions are effective and scalable.Next, we focus on the application aspect and consider in-memory OLAP tools, which have been extensively used by enterprises to make better and faster decisions. Specifically, we take the first attempt towards automatically extracting top-k insights from in-memory OLAP cube. It is useful not only for non-expert users, but also reduces the manual effort of data analysts. It has challenges on (i) the effectiveness of the extracted insights and (ii) the efficiency of top-k insight computation for in-memory data warehouses. We first propose a meaningful scoring function for insights to address (i). Then, we contribute a computation framework for top-k insights, together with a suite of computation optimization techniques to address (ii). Our experimental study on both real data and synthetic data verifies the effectiveness and efficiency of our proposed solution.