Funded Research

The main research approach that we have taken is to gain deep understanding of domain applications' needs and to rethink systems' core abstractions based on that understanding; we iterate between these two activities. We have been motivated by real problems experienced by domain users, developers, and operators in practice. With these first-hand experiences, we perform comprehensive measurements to pinpoint system bottlenecks, characterize application behaviors and their needs, and understand practitioners' pain points.

A systems solution often requires designing a new abstraction or adapting an existing abstraction in a novel way (e.g., in a new scenario, under new settings or assumptions). We take different routes when designing new systems solutions. On the one hand is the freedom of doing blue sky research, where we openly question the fundamentals underlying distributed systems in face of rapidly-emerging needs and challenges. For example, we rethink the design of modern cloud storage systems and build the first tensor-centric AI storage system that exploits tensor-level relationships to reduce storage cost and achieve sustainability. On the other hand, we seek practical and easily-deployable solutions for otherwise sophisticated systems problems. Distributed systems often involve complex cross-component interactions in order to support cross-cutting tasks such as scheduling. This leads us to gravitate towards simple and general solutions that solve not a specific problem, but which satisfy a general set of applications. For instance, we design a new function orchestrator which reduces production FaaS cold start cost and is deployed in production cloud platforms.