Dynamic View Materialization to Improve OLAP Drill-Down Operations

• Main Authors: Chia-Chen Lin, 林家震
• Other Authors: Yungho Leu
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/18113715280091617652

碩士 === 國立臺灣科技大學 === 資訊管理系 === 88 === View materializations can efficiently expedite the execution of OLAP queries. Traditionally, view materializations in Relational OLAP (ROLAP) are performed statically. That is, they are performed by only considering the cost model without taking the user’s query behavior into account. As a result, only the execution time for the materialization plan rather than the execution times of the user’s queries is optimized. In this thesis, we propose several methods that materialize views in a data cube by considering the user’s query behavior. The major data cube operation that is considered is the Drill-Down operation. The reason to choose the Drill-Down operation is that the Drill-Down operation is the most frequently issued and the most useful operation among the Cube operations in an ROLAP. Through collection and analyzing the use’s query histories, we outlined two materialization schemes─the path-oriented method and the segment-oriented method. In the path-oriented method, we keep track the count of the path reference and, then, use the count as the criterion to prioritize the order of materialization of views. While in the segment-oriented method, we keep track the count of the segment reference and use it to prioritize the order of Materialization. Through comprehensive experiments, we show that when the Drill-Down operation is the major user activities, the path-oriented method outperforms the segment-oriented method. On the contrary, when the Drill-Down operation does not prevail in the cube operations, the segment-oriented method performs better than the path-oriented method. We also consider the materialization problem in a hierarchy of a single dimension. In the hierarchy of a dimension, we measure the degree of clustering in each level of the hierarchy. Using the saving cost per unit of memory as the prioritization criterion, we prove that the level with the highest degree of clustering is the one that should be materialized first.