Jun 12

Print this Post

GCOE Talk 13/06/16 | Group By Operator on GPU – continued

In a previous presentation, we presented a GPU Group By Operator, which shows many interesting effects due to unavoidable random memory accesses (Paper [1]). From there, we continued our research in two directions: (1) we ported the operator to OpenCL to compare the performance to CUDA, mainly to see if the effects are architectural and therefore present with both implementations. Additionally, we tested the operator on AMD CPUs and GPUs, Intel CPUs, GPUs and the Xeon Phi, as well as the IBM cell and power processor. In a second direction (2), we looked closer at the most significant effect, a performance drop when accessing more than 2 GB of data. We explain the architectural background to some extent and we actually found a simple software based solution,avoiding the performance problems at all.

The talk will take place at 3:40pm in TRE-101 VC, Zellescher Weg 16, TU Dresden.

[1] Karnagel, T.; Mueller, R.; Lohman, G.; Optimizing GPU-accelerated Group-By and Aggregation. In Proceedings of the VLDB Endowment 2015 (ADMS’15), August 31, 2015, Kohala Coast, Hawai’i, USA

Permanent link to this article: https://gcoe-dresden.de/gcoe-talk-130616-group-by-operator-on-gpu/