Chiplet-based CPUs, which combine multiple independent chips on a single die, allow hardware to scale to higher core counts at the cost of more memory heterogeneity and performance variability. This raises challenges when existing query engines are deployed on chiplet-based CPUs: current designs make assumptions about uniform memory access, cache locality and consistent core performance, which lead to ineffective CPU utilization.

In this paper, we analyse these performance impact when query engines ignore chiplet-specific properties. We show that naive deployment approches result in a significant degradation in query processing efficiency, resulting in non-linear scaling even within a single CPU socket domain. Based on a comprehensive experimentation evaluation, we propose approaches to deploy engines on chiplet-based CPUs with maximum performance. We show that distributing processing tasks according to chiplet topology ensures optimal resource utilization, maximizing performance and scalability. Such a chiplet-conscious deployment strategy of a shared-nothing query engine can enhance performance by up to 7× compared to hardware-oblivious approaches.