A framework for multidimension al indexes on distributed and highly-available data stores

Authors Cesare Cugnasco, Yolanda Becerra, Jordi Torres
Title A framework for multidimension al indexes on distributed and highly-available data stores
Abstract Spatial Big Data is considered an essential trend in future scientific and business applications. Indeed, research instruments, medical devices, and social networks generate hundreds of peta bytes of spatial data per year. However, as many authors have pointed out, the lack of specialized frameworks dealing with such kind of data is limiting possible applications and probably precluding many scientific breakthroughs. In this thesis, we describe three HPC scientific applications, ranging from molecular dynamics, neuroscience analysis, and physics simulations, where we experience first hand the limits of the existing technologies. Thanks to our experience, we define the desirable missing functionalities, and we focus on two features that when combined significantly improve the way scientific data is analyzed. On one side, scientific simulations generate complex datasets where multiple correlated characteristics describe each item. For instance, a particle might have a space position (x,y,z) at a given time (t). If we want to find all elements within the same area and period, we either have to scan the whole dataset, or we must organize the data so that all items in the same space and time are stored together. The second approach is called Multidimensional Indexing (MI), and it uses different techniques to cluster and to organize similar data together. On the other side, approximate analytics has been often indicated as a smart and flexible way to explore large datasets in a short period. Approximate analytics includes a broad family of algorithms which aims to speed up analytical workloads by relaxing the precision of the results within a specific interval of confidence. For instance, if we want to know the average age in a group with 1-year precision, we can consider just a random fraction of all the people, thus reducing the amount of calculation. But if we also want less I/O operations, we need efficient data sampling, which means organizing data in a way that we do not need to scan the whole data set to generate a random sample of it. According to our analysis, combining Multidimensional Indexing with efficient data Sampling (MIS) is a vital missing feature not available in the current distributed data management solutions. This thesis aims to solve such a shortcoming and it provides novel scalable solutions. At first, we describe the existing data management alternatives; then we motivate our preference for NoSQL key-value databases. Secondly, we propose an analytical model to study the influence of data models on the scalability and performance of this kind of distributed database. Thirdly, we use the analytical model to design two novel multidimensional indexes with efficient data sampling: the D8tree and the AOTree. Our first solution, the D8tree, improves state of the art for approximate spatial queries on static and mostly read dataset. Later, we enhanced the data ingestion capability or our approach by introducing the AOTree, an algorithm that enables the query performance of the D8tree even for HPC write-intensive applications. We compared our solution with PostgreSQL and plain storage, and we demonstrate that our proposal has better performance and scalability. Finally, we describe Qbeast, the novel distributed system that implements the D8tree and the AOTree using NoSQL technologies, and we illustrate how Qbeast simplifies the workflow of scientists in various HPC applications providing a scalable and integrated solution for data analysis and management.
Institute Universitat Politècnica de Catalunya. Departament d'Arquitectura de Computadors
Date 22 March 2019
Location Barcelona, Spain
Url http://hdl.handle.net/2117/131429