My research makes computing and networking infrastructure predictable and resilient for AI and real-time interactive applications, from cloud datacenters to the network edge. I build abstractions and runtime systems for predictable performance under resource sharing and failures, along with faithful surrogates of production that guide how they are designed, configured, and tested.
Recent News [all]
- [09/2026] NicOS and Programmable Scale-Up Switches accepted to appear at HotNets 2026
- [09/2026] Dooly accepted to appear at NeurIPS 2026
- [08/2026] Vulcan accepted to appear at EuroSys 2027
- [07/2026] Received an NSF VINES award for FarmBrain, real-time agricultural robotics with AI-native cellular edge computing
- [07/2026] LFS accepted to appear at NSDI 2027
- [05/2026] PacketExpress accepted to appear at SIGCOMM 2026
- [03/2026] Mimesys accepted to appear at OSDI 2026
- [12/2025] SMEC accepted to appear at NSDI 2026
Research Themes
Programmable and Resilient Datacenter Infrastructure for AI and Cloud
NICs, switches, and GPU fabrics have become shared, programmable computers at the core of AI and cloud infrastructure, yet software can neither fully program nor manage them. We build abstractions and runtime systems that make this hardware programmable, manage and isolate its shared resources like an operating system, and keep workloads running through failures.
Leto (arXiv'26) · NicOS (HotNets'26) · Programmable Scale-Up Switches (HotNets'26) · Alkali (NSDI'25) · ExoPlane (NSDI'23) · RedPlane (SIGCOMM'21) · TEA (SIGCOMM'20)
Predictable and Resilient Real-Time Edge Computing
Real-time interactive applications, from live video analytics to live generative AI, need end-to-end latency guarantees across the wireless network and edge servers, yet both are scheduled without awareness of application SLOs. We build systems that make computing at the network edge SLO-aware, secure, and resilient to failures.
SMEC (NSDI'26) · 5G Fronthaul Security (USENIX Security'24) · Slingshot (SIGCOMM'23) · Atlas (MobiCom'23)
Faithful Surrogates for Production
A configuration, placement, or policy is only as good as the testbed it is evaluated in, yet production, the only fully faithful testbed, is too costly to measure and too private to share. We build cheap, faithful surrogates for both what is deployed (hardware, software, and models) and what it is deployed into (traffic and co-located tenants), from data we can actually get, such as operation profiles and resource-usage traces, and through high-fidelity simulation.
Dooly (NeurIPS'26) · Mimesys (OSDI'26) · TraceLLM (EMNLP'25)