Datum und Uhrzeit:
Ort: Large Seminar room OMZ (U014, INF 350, floor -1)
Title: Oasis: Hiding the Cost of Querying Parquet Files in the Datapath
Jonas Dann from ETH Zürich will present the latest research on accelerating cloud-native database systems through hardware offloading.
Abstract: Cloud-native database systems separate compute and storage to improve elasticity and resource utilization. However, querying compressed data stored in data lakes and lakehouses requires decoding file formats such as Parquet for every query, introducing a significant processing overhead. As storage and network bandwidth continue to improve faster than CPU cost-performance, this decoding increasingly limits the cost efficiency of cloud deployments.
This talk presents Oasis, a SmartNIC-based data processing architecture that offloads Parquet decoding into the network datapath using a dedicated hardware accelerator. By overlapping Parquet decoding with query execution, Oasis hides much of the decoding cost from the query engine. Experimental results demonstrate that this approach can almost perfectly overlap decoding and query processing for many workloads. End-to-end, Oasis nearly doubles the query throughput of DuckDB on Parquet files until the PCIe interconnect becomes the performance bottleneck.
CV: Jonas Dann is a doctoral researcher at the Systems Group of ETH Zürich. His research focuses on database systems, cloud infrastructure, and hardware acceleration for data-intensive applications. His work explores how emerging hardware technologies, including SmartNICs and specialized accelerators, can improve the performance and efficiency of cloud-native data processing systems.