In-pipeline parallel execution using pass block buffer in column-stores

Chip multiprocessor(CMP) becomes rapidly the leading role of the processor market due to its low-power dissipation, low cost and some other advantages. It provides hardware support for multithreading. Column-stores have significant advantages in OLAP. In column-stores, utilizing fully multi-core resources can greatly improve performance of query processing. To this end, this paper establishes passing block buffer to make main thread and worker threads read and write different passing blocks respectively, then parent node and child node of the physical execution tree execute in parallel. We use the classic producer-consumer pattern to solve the problem of synchronization between these threads. In the column-store system DWMS developed by our laboratory, experimental results on benchmark data set SSB show the effectiveness of this design, it can improve 50% execution performance for some typical complex queries.