Adaptive Memory Hierarchy Management

现代计算机依靠内存层次结构(由磁盘、主存和硬件缓存组成)来实现高性能和低成本。然而，技术趋势继续增加各层之间的访问时间差距。除非能找到新的技术来提高命中率和数据移动效率，否则这些差距可能会降低内存层次结构的有效性。研究集中在三个这样的技术上：自适应虚拟内存、近最佳页面着色和减少一致性缺失。自适应虚拟内存的目标是磁盘层和主内存层。它通过观察应用程序的页面错误模式并相应地调整页面替换和预取策略，以及通过将动态交换分配与页面管理相集成来提高性能。近乎最佳的页面颜色以主存和硬件缓存层为目标。它通过小心地将页面放入高速缓存集来提高高速缓存命中率。本研究从理论上研究了该问题，并在此基础上设计了包括轮廓优化在内的实用算法。减少多处理器缓存中的粘性缺失。它使用一个数据库系统作为案例研究，以找出对高速缓存一致性未命中贡献最大的常见数据结构和算法，并设计和评估性能更好的备选方案。

Modern computers rely on memory hierarchies (which consist of disks, main-memory and hardware caches) for high performance and low cost. Technology trends, however, have continued to increase the access-time gaps between the layers. The gaps threaten to reduce the effectiveness of memory hierarchies unless novel techniques can be found to improve hit ratios and data movement efficiencies. The research focuses on three such techniques: adaptive virtual memory, near-optimal page coloring, and coherence- miss reduction. Adaptive virtual memory targets the disk and main- memory layers. It improves performance by observing applications' page fault patterns and adjusting page replacement and prefetching policies accordingly, and by integrating dynamic swap allocation with page managements. Near-optimal page coloring targets the main- memory and hardware cache layers. It increases cache hit ratio by carefully placing the pages into cache sets. The research studies the theoretical aspects of the problem and designs practical algorithms including profile-based optimizations based on the theoretical foundation.Coherence-miss reduction targets multi- processor caches. It uses a database system as a case study to identify common data structures and algorithms that contribute the most to cache coherence misses, and design and evaluate alternatives that perform better.Combined, the three techniques promise to improve memory hierarchy performance significantly.