Xudong Liao
PhD Candidate, HKUST, Hong Kong SAR, China

I am a Ph.D candidate in Hong Kong University of Science and Technology (HKUST), advised by Prof. Kai Chen. Before that, I received my B.Eng in Software Engineering in Wuhan University (Outstanding Graduate) in 2020.
In my research projects, I focus on:
- application-aware optimizations for distributed systems. I build systems that push the hardware–software boundary by tailoring architectures and algorithms to each workload’s unique behavior:
- MixNet (SIGCOMM’25) – a runtime reconfigurable optical–electrical fabric that leverages the dynamic, sparse and localized traffic patterns in distributed Mixture-of-Experts training to regionally adapt its topology on-the-fly, enabling scalable and cost-efficient training across thousands of GPUs—with near-ideal training speed and significantly reduced networking costs.
- Herald (NSDI’24) – an embedding-aware scheduler that exploits the predictable and infrequent in-cache embedding access patterns of DLRM training to schedule the embedding access, eliminating a substantial portion of communication and efficiently speeding up the training.
- Pallas (ATC’25) – an rack-scale CPU scheduling system that utilizes switch programmability and request-level predictability to enable efficient in-network workload shaping, driving near-optimal microsecond-level tail latency.
- designing performant and practical DRL-driven CC algorithms – including MOCC (EuroSys’22), Spine (CoNEXT’22), Astraea (EuroSys’24), and Jury (EuroSys’25). These projects tackle key obstacles such as multi-objectives, overhead, fairness & convergence, and performance generalizability, paving the way for real-world deployment of DRL-based transport protocols.
I was fortunate to be advised by Prof. Yanjiao Chen during my time at WHU. Additionally, I am fortunate to collaborate closely with Prof. Guyue Liu from Peking University and Dr. Zhizhen Zhong from MIT on several recent projects.
Research Interests
- Machine Learning System
- Datacenter Networking
- Congestion Control
- Optical Network
news
Aug 22, 2025 | Two co-authored papers LTP and MFS accepted to EuroSys 2026! |
---|---|
Jul 12, 2025 | MixNet accepted to SIGCOMM 2025! |
Apr 25, 2025 | Pallas accepted to ATC 2025! |
Jan 10, 2024 | Astraea accepted to EuroSys 2024! |
Dec 07, 2023 | Herald accepted to NSDI 2024! |
selected publications
* equal contribution
View Full Publication List →- ATCTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload ShapingIn 2025 USENIX Annual Technical Conference (ATC 2025) , 2025
- OSDIEnabling Efficient GPU Communication over Multiple NICs with FuseLinkIn Proceedings of the 19th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2025) , 2025
- EuroSysAchieving Fairness Generalizability for Learning-based Congestion Control with JuryIn Proceedings of the 20th ACM European Conference on Computer Systems (EuroSys 2025) , 2025