Providing Timely Results with an Elastic Parallel DW

Abstract

OLAP analysis is a fundamental tool for enterprises in competitive markets. While known (planned) queries can be tuned to provide fast answers, ad-hoc queries have to process huge volumes of the base DW data and thus resulting in slower response times. While parallel architectures can provide improved performance, by using a divide-and-conquer approach, their structure is rigid and suffers from scalability limitations imposed by the star schema model used in most deployments. Therefore usually they are over-dimensioned with computational resources in order to provide fast response times. However, for most business decisions, it is more important to have guarantees that queries will be answered in a timely fashion. The star schema model physical representation introduces severe limitations to scalability and in the ability to provide timely execution, due to the well-known parallel join issue and the 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. In this paper, we propose PH-ONE an architecture that overcomes the scalability limitations by combining an elastic set of inexpensive heterogeneous nodes with a denormalized DW storage model organization, which requires a minimal set of predictable processing tasks, using in a shared-nothing scheme to remove costly joins. PH-ONE delivers timely execution guarantees by adjusting the number of processing nodes and by rebalancing the data load according to the nodes characteristics. We used the TPC-H benchmark to evaluate PH-ONE ability to provide timely results.