讲座题目：容量约束下的GPU内存自动管理框架 主讲人：杨峻 教授 开始时间：2019-08-12 10:30:00 结束时间：2019-08-12 11:30:00 讲座地址：华东师范大学数学馆201 主办单位：计算机科学与技术学院
报告人简介： 杨峻，匹兹堡大学电子与计算机工程系教授。研究方向为计算机体系结构，研究方向包括GPU设计、新兴的内存技术、互连网络、3D集成以及电源和功耗管理技术。获得了2008年的NSF CAREER award、2010年的IEEE Micro Top Picks奖、2013年ISLPED和2007年ICCD最佳论文奖。被列入了HPCA名人堂。 报告内容： Memory capacity in GPUs has been a major challenge for today's data-intensive workloads. Traditionally, a programmer needs to manually divide a workload and its data to fit it into the limited GPU memory space. Unified Virtual Memory (UVM) was developed to support on-demand paging and data migration, which dramatically reduces developer effort. However, such support result in great performance loss during the automatic data movement. In this talk, I will introduce a memory management framework, called ETC, that transparently improves GPU performance under memory oversubscription using new techniques to overlap eviction latency, reduce thrashing cost, and increase effective memory capacity. We develop a tree-based eviction policy (E) that coordinates with hardware prefetching semantics to maintain memory locality during data movement. Next, memory thrashing can be ameliorated with memory-aware throttling (T), which dynamically reduces the GPU parallelism when page fault frequency becomes high. Finally, capacity compression (C) can enable larger working sets without increasing physical memory capacity. No single technique fits all workloads, and, thus ETC integrates the above techniques into a principled framework that dynamically selects the most effective combination of techniques, transparently to the running software. Our evaluation shows that ETC fully mitigates the oversubscription overhead for regular applications and outperform the state-of-the-art baseline for other applications with specific data sharing properties. |
