8月12日 杨峻:容量约束下的GPU内存自动管理框架

时间:2019-08-04浏览:132设置


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 award2010年的IEEE Micro Top Picks奖、2013ISLPED2007ICCD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.

  


返回原图
/