SHF:Small: Optimization of Parallel and Shared Cache Memory using the Footprint Theory

多核处理器为个人、科学和商业计算平台带来了计算能力的巨大增长。这些处理器中的大多数片上内存都用于共享缓存，使其成为性能的主要因素。 一种常见的做法是通过测试构建改进路径来提高缓存性能。 然而，测试可能需要太多的步骤，它可能不会收敛到一个稳定的解决方案，最重要的是，对于大规模的并行程序，解决方案是远远不是最佳的，因为测试只涵盖了所有可能性的一小部分。 这项研究为软件开发人员和硬件工程师提供了基于建模的新工具，而不是测试。 该研究建立在过去的NSF支持的研究，开发了足迹理论的顺序应用程序。该工作解决了数据共享的新问题，并扩展了局部性理论，以优化并行代码。并行系统需要复杂的模型，这种复杂性是由可组合的模型来解决大规模建模的大规模问题。新工具包括程序和机器特性的统计模型。 程序模型包括剖析分析，该剖析分析在所有线程组合和数据放置中导出数据和高速缓存共享，在所有大小的高速缓存中。高速缓存模型分析了许多硬件设计的影响，包括关联性，排他性，一致性，高速缓存方式分区和瞬态加载/存储。这些模型的组合使用，使并行程序优化，并改善线程和数据放置。

Multicore processors bring a tremendous increase in computing power to personal, scientific, and business computing platforms. Most on-chip memory in these processors is devoted to shared cache, making it a major factor in performance. A common practice is to improve cache performance by building a path of improvements through testing. However, testing may take too many steps, it may not converge to a stable solution, and most importantly for large scale parallel programs, the solution is far from optimal since testing covers only a minuscule fraction of all possibilities. Instead of testing, this research provides software developers and hardware engineers new tools based on modeling. The research builds on the past NSF supported research which has developed the footprint theory for sequential applications. This work solves the new problems of data sharing and extends the locality theory to optimize the parallel code. Parallel systems require complex models, and this complexity is addressed by composable models to solve large scale problems with large scale modeling. The new tools include statistical models of both program and machine characteristics. Program models include profiling analysis that derive the data and cache sharing in all thread combinations and data placements, in cache of all sizes. Cache models analyze the effect of many hardware designs including associativity, exclusivity, coherence, cache-way partitioning, and transient loads/stores. Combined use of these models enables parallel program optimization and improves thread and data placement.