一句话：针对 GPU 统一内存工作集超出物理容量导致的数据搬移与性能损失，提出内存超额订阅管理框架。
现代独立 GPU 支持统一内存与按需分页，可在 CPU 内存和 GPU 内存之间自动搬移数据，显著减少开发者显式管理数据的负担。但当应用程序的工作集超过物理内存容量时，这种自动数据搬移会带来大量缺页与传输开销，造成严重性能损失，即内存超额订阅问题。该论文面向这一场景提出一个 GPU 内存超额订阅管理框架，旨在在工作集超出显存容量时有效管理数据放置与迁移，缓解因反复传输和缺页导致的性能下降。摘要未给出具体方法细节与实验数据，标题表明其核心贡献是提供系统化的超额订阅管理机制。
Modern discrete GPUs support unified memory and demand paging. Automatic management of data movement between CPU memory and GPU memory dramatically reduces developer effort. However, when application working sets exceed physical memory capacity, the resulting data movement can cause great performance loss.
@inproceedings{paperbot4937,
title = {A Framework for Memory Oversubscription Management in Graphics Processing Units},
author = {Chen Li and Rachata Ausavarungnirun and Christopher J. Rossbach and Youtao Zhang and Onur Mutlu and Yang Guo and Jun Yang},
booktitle = {ASPLOS},
year = {2019},
doi = {10.1145/3297858.3304044}
}