Statusvortrag von Herrn Andreas Geyer

Title: Workload-Aware Interconnect Optimization in Analytical Data Engines

Abstract: While monolithic architectures remain the standard for database systems, memory disaggregation becomes more prevalent to address resource inefficiencies and rigid scaling limits. However, physically separating compute and memory introduces new hardware constraints for database management systems: remote interconnects exhibit lower bandwidth and higher latency compared to local memory. This potential bottleneck is particularly noticeable in analytical workloads, where concurrent queries often redundantly fetch overlapping working sets across the remote link.

Consequently, this architectural shift requires a re-evaluation of established query execution and optimization paradigms to determine their viability under the constraints of remote memory. In this talk, I will examine the limitations of current execution techniques in disaggregated setups and introduce Pipeline Grouping. By dynamically synchronizing remote data access at runtime, Pipeline Grouping reduces redundant I/O, aiming to better align query execution with the physical hardware limits of disaggregated database architectures.

Betreuer: Prof. Wolfgang Lehner