Optimization and Execution of Complex Scientific Queries over Uncorrelated Experimental Data

摘要

Scientific experiments produce large volumes of data represented as complex objects that describe independent events such as particle collisions. Scientific analyses can be expressed as queries selecting objects that satisfy complex local conditions over properties of each object. The conditions include joins, aggregate functions, and numerical computations. Traditional query processing where data is loaded into a database does not perform well, since it takes time and space to load and index data. Therefore, we developed SQISLE to efficiently process in one pass large queries selecting complex objects from sources. Our contributions include runtime query optimization strategies, which during query execution collect runtime query statistics, reoptimize the query using collected statistics, and dynamically switch optimization strategies. Furthermore, performance is improved by query rewrites, temporary view materializations, and compile time evaluation of query fragments. We demonstrate that queries in SQISLE perform close to hard-coded C++ implementations of the same analyses.