GPU and coprocessor use in analytic query processing - Why we have only just begun

The past several years have seen several initial efforts to speed up analytic DB query processing using coprocessors (GPUs mostly) [1]. But DBMSes are complex software systems, which have seen decades of spirited evolution and optimization on CPUs - and coprocessor proponents have found it very challenging to catch up. Thus, only last year was a system presented [2] which surpasses MonetDB-level performance on TPC-H queries. Yet, that system is still slow compared to the CPU state-of-the-art; and it has remained closed and unreleased - exemplifying two aspects of the challenges of putting coprocessors to use: The technical and the social/methodological.

Drawing inspiration from shortcomings of existing work (with GPUs), and from both technical and social aspects of leading projects (HyPer, VectorWise and MonetDB in its own way), we will lay out some of these challenges, none having been seriously tackled so far; argue for certain approaches (GPU-specific and otherwise) for addressing them; and if time allows, discuss potential benefits from the interplay of such approaches.

[1] Breß, Sebastian, et al. "GNU-accelerated database systems: Survey and open challenges." Transactions on Large-Scale Data-and Knowledge-Centered Systems XV. Springer Berlin Heidelberg, 2014. 1-35.

[2] Overtaking CPU DBMSes with a GPU in Whole-Query Analytic Processing with Parallelism-Friendly Execution Plan Optimization, Adnan Agbaria, David Minor, Natan Peterfreund, Eyal Rozenberg, and Ofer Rosenberg.