Optimizing select conditions on GPUs

Abstract

Implementations of data processing operators on GPU processors have achieved significant performance improvements over their multicore CPU counterparts. To achieve maximum performance, database operator implementations must take into consideration special features of GPU architectures. A crucial di↵erence is that the unit of execution is a group ("warp") of threads, 32 threads in our target architecture, as opposed to a single thread for CPUs. In the presence of branches, threads in a warp have to follow the same execution path; if some threads diverge then di↵erent paths are serialized. Additionally, similarly to CPUs, branches degrade the e ciency of instruction scheduling. Here, we study conjunctive selection queries where branching hurts performance. We compute the optimal execution plan for a conjunctive query, taking branch penalties into account and consider both single-kernel and multi-kernel plans. Our evaluation suggests that divergence a↵ects performance significantly and that our techniques reduce resource underutilization and improve operator performance.

Accelerating In-Memory Database Selections Using Latency Masking Hardware Threads

“…We hold the same assumption for the CPU and the GPU. We do not consider any transfer time for all the devices, which is usually the case for in-memory data analytics [5,35,70].…”

“…Our GPU implementation relies on the findings of References [36,70]. We assign each GPU thread to a distinct tuple for processing and evaluate the conditions of a query using a for-loop.…”

“…Each evaluation is aggregated to a local register using bitwise-AND operands. Also, as suggested in Reference [70], we use optimal GPU plans to mitigate the effects of thread divergence on GPU. The final evaluation results are written back to global memory into a flag vector.…”

“…This grouping is not only relevant to computation but also to global memory accesses. In fact, the related literature has unanimously promoted the use of column major [11,24,31,70] data layouts to maximize the effective memory bandwidth. Leveraging the aforementioned data layout, database researchers focused on defining a set of fairly standard parallel primitives [36] having a direct correlation to relational operators.…”

“…For example, workloads from complex database analytics, hash joins [35], data mining, graph algorithms, bioinformatics, and so on exhibit a very poor degree of spatial and/or temporal localities and do not benefit from large cache hierarchies. (2) Control flow irregularity is caused by the dynamic control flow and leads to branch mis-prediction; this contributes to a large fraction of stall time on CPUs [62,74] or thread divergence on GPUs [70]. An example that exhibits such behavior is the selection operator.…”