A Tale of Two Paths - Hybrid Data Plane for Efficient Far-Memory Applications

Explore a groundbreaking conference talk from OSDI '24 that delves into the development of Atlas, a hybrid data plane for efficient far-memory applications. Learn about the two primary data paths in existing far memory systems: kernel-based paging and object-level fetching. Discover how Atlas combines these approaches through runtime-kernel co-design to optimize performance for real-world applications. Understand the innovative use of always-on profiling to measure page locality and dynamically switch between paging and object fetching. Gain insights into how Atlas improves throughput and reduces tail latency compared to state-of-the-art techniques like AIFM and Fastswap. Dive into the technical details of breaking the memory capacity wall and enhancing efficiency in far memory systems during this 16-minute presentation by researchers from the University of Chinese Academy of Sciences, UCLA, Tsinghua University, and Microsoft Research.