TEEPA: a timely-aware elastic parallel architecture

Abstract

Parallel Shared-Nothing architectures are frequently used to handle large star-schema Data Warehouses (DW). The continuous increase in data volume and the star-schema storage organization introduce severe limitations to scalability due to the well-known parallel join issues and the resulting need to use solutions such as on-the fly repartitioning of data or intermediate results, or massive replication of large data sets that still need to be joined locally, constraining their ability to deliver fast results. Parallelism may improve query performance, however some business decisions may require that query results be timely available which, even with additional parallelism and significant upgrade costs (both monetary and due to disturbance of normal operations), cannot be guaranteed. We propose a Timely-aware Execution Parallel Architecture (TEEPA) which balances data load and query processing among an elastic set of non-dedicated heterogeneous nodes in order to provide scale-out performance and timely query results. Data is allocated using adaptable storage models to minimize join costs (the major uncertainty factor) which best fit the nodes' capabilities, while preserving a consistent logical view of the star-schema. We present experimental evaluation of TEEPA and demonstrate its ability to provide timely results.