Projects
With Yi Chen, Site Fan, Po-Yu Hou, Maizhe Zhang, and Peng Huang. LiteLib studies how to contain the impact of failures in stateful applications using compact replicas: smaller, cheaper standbys that preserve enough state to recover quickly without the cost of full replication. The project explores replica design and recovery mechanisms that keep applications available when primary replicas fail, and is accepted to appear at NSDI 2027.
With Reto Achermann and Andrea Lattuada, at MPI-SWS and the University of Michigan. Verifying low-level system code requires an abstract hardware model, but the result is only as trustworthy as that model. Validating a formal model against real hardware is difficult because observable behavior depends on many invisible microarchitectural steps. This project develops a more tractable approach to checking x86 memory-management-unit models against real machines, and led to our PLOS 2026 paper Towards Tractable Hardware Model Validation Against Real Hardware.
With Ryan Huang and Mosharaf Chowdhury at the University of Michigan. Modern applications increasingly embed AI features, but operating systems still treat AI runtimes as ordinary user processes. This project explores OS-level management of multiple AI runtimes at once, so concurrent applications can share accelerators, memory, and scheduling policies more safely and efficiently when AI workloads interfere with one another.
With Zixu Tian and Mohan Gurusamy at the National University of Singapore. 5G network slicing expands the attack surface of the distributed user plane, where univariate KPI monitors often miss context-dependent anomalies. We built an extensible simulated 5G / SDN testbed and a dual-branch Transformer that jointly models temporal dependencies and inter-metric correlations, with cross-attention fusion and patch-wise masking for robust reconstruction under partial observability. Experiments on real network traffic show effective detection of diverse user-plane anomalies; the work appeared at IEEE CNS 2025.