Collaborative Research: Tuning Libraries to Effectively Exploit the Memory Hierarchy

我们建议使用三种不同的优化方法来提高库代码在多级内存层次结构上运行的性能。我们将针对阻塞和预处理方法来增加稀疏和结构化矩阵中引用的局部性。接下来，我们使用缓存敏感的数据结构来进一步提高内存层次结构的利用率。最后，我们执行代码和数据布局转换，有效地将库重新映射到底层内存硬件。我们最初的目标应用包括各种Krylov迭代求解器。当我们计划生成这些库代码的优化版本时，我们还将追求库的其他类。我们的优化将解决将代码有效迁移到新硬件平台的需求，并利用新平台上提供的新架构特性。

We propose to improve the performance of library codes as run on a multi-level memory hierarchy using three different approaches to optimization. We will target blocking and preconditioning methods to increase the locality of reference in sparse and structured matrices. Next we employ cache-conscious data structures to further improve memory hierarchy utilization. Finally, we perform code and data layout transformations that effectively remap library to the underlying memory hardware. Our initial target applications include a variety of Krylov iterative solvers. While we plan to produce optimized versions of these library codes, we will also pursue additional classes of libraries. Our optimizations will address the need to effectively migrate code to new hardware platforms and exploit the novel architecture features provided on the new platform.