The Case for SIMDified Analytical Query Processing on GPUs

Data-level parallelism (DLP) is a heavily used hardware-driven parallelization technique to optimize the analytical query processing, especially in in-memory column stores. This kind of parallelism is characterized by executing essentially the same operation on different data elements simultaneously. Besides Single Instruction Multiple Data (SIMD) extensions on common x86-processors, GPUs also provide DLP but with a different execution model called Single Instruction Multiple Threads (SIMT), where multiple scalar threads are executed in a SIMD manner. Unfortunately, a complete GPU-specific implementation of all query operators has to be set up, since the state of the vectorized implementations cannot be ported from x86-processors to GPUs right now. To avoid this implementation effort, we present our vision to virtualize GPUs as virtual vector engines with software-defined SIMD instructions and to specialize hardware-oblivious vectorized operators to GPUs using our Template Vector Library (TVL) in this paper.