Research
Our group works on computer systems for machine learning. We apply a broad set of systems techniques across the stack, from storage and networking up through runtimes and serving frameworks, to make end-to-end machine learning pipelines more performant, scalable, efficient, and secure.
Machine Learning Data Infrastructure
We build storage and data systems to manage the massive amounts of data needed to train and serve models. These systems ensure that the GPUs and accelerators that train and serve models are not starved for data. Our work has been deployed at production scale at Meta.
Representative work: cedar (VLDB'25), Tectonic-Shift (ATC'23), RecD (MLSys'23), Data Preprocessing (ISCA'22)
Training Systems at Scale
We build systems that scale to train massive foundation models more efficiently and reliably. At datacenter scale, training a model involves thousands of GPUs. Our work builds systems that compose these GPUs into a reliable, efficient, and performant training cluster.
Representative work: SYMI (NSDI'26), ReCycle (SOSP'24), ZeroNIC (OSDI'24)
Inference and Agentic Serving
We build serving and caching systems to make agentic applications faster and cheaper. These applications require complex reasoning pipelines with many model calls and large intermediate context. We build serving frameworks that efficiently execute and manage these pipelines.
Representative work: Strata (OSDI'26), Beyond the Accelerator (IEEE Micro), vCache (ICLR'26), Llama (SoCC'21)
Systems for Robotics and the Edge
We build systems to deploy models on the robots and edge devices under tight latency, resource, and power constraints. Our work enables large, state-of-the-art models to run on these embodied devices, enabling them to perform complex tasks.
Representative work: SHARD (EuroMLSys'26)