Rethinking Cloud Data Processing with DPUs
17 Nov 2026
10.30 AM - 11.30 AM
LT5
Jiasheng (Jason) Hu
PhD student
Computer Science at the University of Toronto
Current Students, Industry/Academic Partners
Abstract:
As cloud data systems increasingly disaggregate computation, storage, and memory, data movement across the network has become an increasingly important part of their execution cost. While networks and storage devices continue to grow faster, host-side processing along the I/O path can still consume substantial CPU resources and limit end-to-end performance. Data Processing Units (DPUs), programmable network devices with onboard general-compute processors, memory, and accelerators, offer a new opportunity to rethink where and how this work is performed.
In this talk, I will introduce our vision for data processing with DPUs and explore the opportunities and challenges of leveraging them to improve the efficiency of cloud data systems. Drawing on our research in characterizing DPU performance and designing DPU-accelerated systems, I will discuss how to effectively exploit their diverse hardware capabilities, overcome resource constraints, and offload data processing and I/O operations from host CPUs. I will also highlight the importance of appropriate programming abstractions and careful end-to-end system design in realizing the potential of DPUs.
Speaker:
Jiasheng (Jason) Hu is a PhD student in Computer Science at the University of Toronto, advised by Prof. Qizhen Zhang. His research lies at the intersection of database and networked systems, with a particular interest in hardware offloading and acceleration. His current work explores how SmartNICs and Data Processing Units (DPUs) can improve the performance and efficiency of data systems. His research has been published at conferences including VLDB, CIDR, NSDI, and SoCC. He was also a research intern at Microsoft Research in Redmond, where he worked on optimizing PostgreSQL's asynchronous I/O, resulting in a paper at the VLDB 2026 Industrial Track.