Efficient OLAP Query Processing in Distributed Data Warehouses

Abstract

The success of Internet applications has led to an explosive growth in the demand for bandwidth from ISPs. Managing an IP network requires collecting and analyzing network data, such as flow-level traffic statistics. Such analyses can typically be expressed as OLAP queries, e.g., correlated aggregate queries and data cubes. Current day OLAP tools for this task assume the availability of the data in a centralized data warehouse. However, the inherently distributed nature of data collection and the huge amount of data extracted at each collection point make it impractical to gather all data at a centralized site. One solution is to maintain a distributed data warehouse, consisting of local data warehouses at each collection point and a coordinator site, with most of the processing being performed at the local sites. In this paper, we consider the problem of efficient evaluation of OLAP queries over a distributed data warehouse. We have developed the Skalla system for this task. Skalla translates OLAP queries, specified as certain algebraic expressions, into distributed evaluation plans which are shipped to individual sites. Salient properties of our approach are that only partial results are shipped - never parts of the detail data. We propose a variety of optimizations to minimize both the synchronization traffic and the local processing done at each site. We finally present an experimental study based on TPC(R) data. Our results demonstrate the scalability of our techniques and quantify the performance benefits of the optimization techniques that have gone into the Skalla system.