Workload and hardware storage optimisation for time series processing

Abstract

A massive amount of historical data is available in every domain and application. This has led the OLAP community into proposing a number of cube algebras, although a standard such as the one for the relational algebra has still not been established. In this study, the user-centric Cube Algebra query language is implemented as an extension to the industry-standard PostgreSQL DBMS. This language is a recent addition to the cube algebras and the original work of its authors does not address translation, execution and optimization. Thus, our work addresses such topics in the context of the relational model as the encoding to this OLAP abstraction. For the evaluation of our work, we present a number of time series case studies: (a) querying using Parametric Outlier Detection; (b) querying using a Simple Moving Average; and (c) querying for similarity using Dynamic Time Warping. Our optimization make use of basic query rewriting techniques inspired by relational algebra. Results show a dramatic improvement when applying query rewriting in all cases, ranging from a 4 to 20 times speed-up, and such techniques are not limited to the relational model. Using parallelism, a speed-up is achievable to help efficient data retrieval further, however it is not linear. A comparison with SQL implementations indicates that there is still a reasonable cost for using the OLAP abstraction in most cases. Thus, future research can focus on addressing this cost, particularly when introducing new operators to the algebra.