Query parallelism improves serial query execution performance by orders of magnitude. Getting optimal performance from an already parallelized query plan is however difficult due to its dependency on run time factors such as correct operator scheduling, memory pressure, disk io performance, and operating system noise. Identifying the exact problems in a parallel query execution is difficult due to inter-dependence of these factors. In this paper we present Tomograph, a tool to visualize the parallel query execution performance bottlenecks. Tomograph provides a time ordered view of operator execution aligned with cpu, memory, and disk IO usage, in an operator at a time execution model. We discuss the usage of Tomograph to identify parallelism issues such as low multi-core utilization, erroneous operator scheduling, incorrect data partitioning, and blocking operators. We share our experiences, insights gained and discuss possible solutions to the identified problems.
, ,
Commit: Time Trails (P019)
International Workshop on Testing Database Systems
Database Architectures

Gawade, M., & Kersten, M. (2013). Tomograph: Highlighting query parallelism in a multi-core system. In SIGMOD Record. A.C.M. doi:10.1145/2479440.2479444