Analytical Query Processing Using Heterogeneous SIMD Instruction Sets

Numerous applications gather increasing amounts of data, which have to be managed and queried. Different hardware developments help to meet this challenge. The grow-ing capacity of main memory enables database systems to keep all their data in memory. Additionally, the hardware landscape is becoming more diverse. A plethora of homo-geneous and heterogeneous co-processors is available, where heterogeneity refers not only to a different computing power, but also to different instruction set architectures. For instance, modern Intel® CPUs offer different instruction sets supporting the Single Instruction Multiple Data (SIMD) paradigm, e.g. SSE, AVX, and AVX512. Database systems have started to exploit SIMD to increase performance. However, this is still a challenging task, because existing algorithms were mainly developed for scalar processing and because there is a huge variety of different instruction sets, which were never standardized and have no unified interface. This requires to completely rewrite the source code for porting a system to another hardware architecture, even if those archi-tectures are not fundamentally different and designed by the same company. Moreover, operations on large registers, which are the core principle of SIMD processing, behave counter-intuitively in several cases. This is especially true for analytical query process-ing, where different memory access patterns and data dependencies caused by the com-pression of data, challenge the limits of the SIMD principle. Finally, there are physical constraints to the use of such instructions affecting the CPU frequency scaling, which is further influenced by the use of multiple cores. This is because the supply power of a CPU is limited, such that not all transistors can be powered at the same time. Hence, there is a complex relationship between performance and power, and therefore also between performance and energy consumption. This thesis addresses the specific challenges, which are introduced by the application of SIMD in general, and the heterogeneity of SIMD ISAs in particular. Hence, the goal of this thesis is to exploit the potential of heterogeneous SIMD ISAs for increasing the performance as well as the energy-efficiency