Research
My research is in cloud-native and distributed data processing systems, with a focus on network communication and emerging data center hardware. I study how DPUs (SmartNICs) can help data systems move and process data more efficiently, and how to design practical abstractions and optimizations for heterogeneous hardware.
Publications
Data processing with DPUs
Making Sense of DPU Performance for Cloud Data Processing — SoCC 2026.
Jiasheng Hu, Chihan Cui, Yuanfan Chen, Philip A. Bernstein, Jialin Li, Qizhen Zhang.
We benchmark recent production DPUs to characterize their performance and investigate where DPU offloading can benefit database workloads.
dpKernels: Harvesting DPU Compute Resources for Data-path Efficiency in Cloud Data Processing — VLDB 2026. Code.
Jiasheng Hu, Kaiwen Zheng, Anna Li, Sidharth Sankhe, Philip A. Bernstein, Qizhen Zhang.
We introduce dpKernels, a set of portable primitives for accessing DPU accelerators and SoC cores, and dpManager, which manages and schedules those resources across DPU platforms.
PD3: Prefetching Data with DPUs for Disaggregated Memory — NSDI 2026. Code.
Sidharth Sankhe, Felix Zhang, Umayrah Chonee, Sherman Lim, Jiasheng Hu, Jialin Li, Qizhen Zhang.
PD3 uses DPUs and application information to prefetch data from disaggregated memory, avoiding compute-server cache misses and offloading network and DMA work from the host.
DPDPU: Data Processing with DPUs — CIDR 2025.
Jiasheng Hu, Philip A. Bernstein, Jialin Li, Qizhen Zhang.
This vision paper outlines opportunities and challenges in using DPUs for data processing and proposes a framework for organizing DPU-enabled system designs.
DDS: DPU-optimized Disaggregated Storage — VLDB 2024. Code.
Qizhen Zhang, Philip Bernstein, Badrish Chandramouli, Jiasheng Hu, Yiming Zheng.
DDS uses a DPU to process disaggregated-storage requests, combining DMA, zero-copy, and userspace I/O to reduce latency and host CPU use with minimal DBMS changes.
Database storage and I/O
Tuning the Lookahead Distance for PostgreSQL Asynchronous IO — VLDB 2026 Industrial Track.
Wentao Wu, Jiasheng Hu, Manoj Syamala, Andres Freund, Vivek Narasayya.
We adaptively tune the lookahead distance used by asynchronous database reads, using I/O completion-time feedback to improve prefetching decisions.