This session explores a novel Linux kernel-space Serverless framework designed to bridge the gap between physically disaggregated hardware pools (e.g. disaggregated memory) and the instant execution semantics of modern Serverless functions. By building a three-layer architecture spanning an RDMA network foundation, distributed OS kernel primitives (RDMA Fork, Map, and mmap), and application-specific scaling strategies, we demonstrate how to achieve millisecond-level elasticity across disaggregated resources. Attendees will dive into three practical cases, including rapid GPU scale-out for LLM inference, dynamic memory scale-up for recommendation systems, and decoupled I/O capacity scaling for storage engines.

Speakers:

Mingxuan Liu: Postdoctoral Fellow, University of Macau

Mingxuan Liu is a Postdoctoral Fellow at the University of Macau. With over a decade of research experience, his work spans operating system kernels, RDMA networking, serverless computing, and LLM infrastructure. His doctoral research focused on kernel-level serverless resource elasticity for RDMA-based disaggregated architectures. He has published more than ten first-author papers in leading CCF-A/B-ranked academic venues. At the University of Macau, he continues his research on high-performance distributed systems.