Toward GPU-accelerated Database Optimization

Abstract

For over three decades, research investigates optimization options in DBMSs. Nowadays, the hardware used in DBMSs become more and more heterogeneous, because processors are bound by a fixed energy budget leading to increased parallelism. Existing optimization approaches in DBMSs do not exploit parallelism for a single optimization task and, hence, can only benefit from the parallelism offered by current hardware by batch-processing multiple optimization tasks.Since a large optimization space often allows us to process sub-spaces in parallel, we expect large gains in result quality for optimization approaches in DBMSs and, hence, performance for query processing on modern (co-)processors. However, parallel optimization on CPUs is likely to slow down query processing, because DBMSs can fully exploit the CPUs computing resources due to high data parallelism. In contrast, the communication overhead of co-processors such as GPUs typically lead to plenty of compute resources unused.In this paper, we motivate the use of parallel co-processors for optimization in DBMSs, identify optimization problems benefiting from parallelization, and show how we can design parallel optimization approaches on the example of the operator placement problem.