Experiments on Static Data Summarization Techniques

The persistent usage of relational data leads to a huge data growth. This immense growth requires faster access and analysis of data. This can be done by either fragmenting the data or by scanning less amount of data as in data blocking and skipping. Data Summarization describes data in a compact way, such that when a query gets executed it only needs to scan the summary of the block and not the entire block to identify which block to scan. Some of the static data summarization techniques such as MinMax, Global Range Encoding (GRE), and Fingerprint Headers are evaluated in this work. The experiment is performed on a benchmark TPC-H dataset, which is a business-oriented dataset. The identified evaluation parameters are: Query Excecution Time (QET) and Percentage of Tuples Skipped. The evaluation is further investigated in terms of Algorithm Execution Time (AET). The QET of Fingerprint Headers is less compared to other techniques as these techniques scans less data compared to the rest of techniques. AET for data sizes x, 1.5x, and 2x is recorded and it is observed that GRE, and two-level Fingerprint Headers records highest AET. This is due to the number of partitions generated by these techniques. Two-level Fingerprint Headers skip the highest number of tuples i.e. 76% as compared to Full Scan as these techniques scan less amount of data.