Selection on Modern CPUs

Abstract

Modern processors employ sophisticated techniques such as speculative or out-of-order execution to hide memory latencies and keep their pipelines fully utilized. However, these techniques introduce high complexity and variance to query processing. In particular, these techniques are transparent to DBMS operations since they are managed by processors internally. To fully utilize the sophisticated capabilities of modern CPUs, it is necessary to understand their characteristics and adjust operators as well as cost models accordingly. In this paper, we extensively examine the execution of a relational selection operator on modern hardware in an in-depth performance analysis. We show, that branching behavior and memory exploitation are two main contributors to run-time. Based on these insights, we show how two common cost models would predict execution costs and why they fall short in determining run-time behavior for parallel execution. We reveal, that cost models which exploit only one performance parameter to determine execution costs are not able to predict the non-linear performance characteristics of modern CPUs.