Compiling for Locality in Advanced Scientific Computations

在现代处理器体系结构中，程序只有在拥有足够的数据局部性来利用片上缓存的情况下才能获得良好的性能。这项研究的重点是开发和评估软件支持，以改善用于顺序和并行机器的高级科学应用程序的局部性。基本前提是编译时分析和复杂的运行时系统都是必需的。运行时系统是必需的，因为许多程序不能静态分析。编译器支持对于向运行时系统插入接口和在可能的情况下直接应用程序转换都是至关重要的。该建议研究了高级科学应用程序中发现的三个特性(3D数组、不规则访问和指针)所需的局部性优化。它专注于扩展局部性优化，以处理多数据、深层内存层次结构(多级缓存和TSB)、硬件和软件预取、非线性内存布局、并行和集群体系结构以及内存性能工具之间的缓存冲突。局部性优化将应用于具有代表性的程序，并在先进的计算机系统上进行实验评估。这项研究的结果将有助于提高计算密集型科学应用的性能。由于计算机体系结构的趋势，吸取的经验教训也同样适用于图像处理和高性能数据库等应用领域。

With modern processor architectures, programs can achieve goodperformance only if they possess sufficient data locality to exploiton-chip caches. This research focuses on developing and evaluatingsoftware support for improving locality for advanced scientificapplications for both sequential and parallel machines. The basicpremise is that both compile-time analyses and sophisticated run-timesystems are necessary. Run-time systems are needed because manyprograms are not analyzable statically. Compiler support is crucialboth for inserting interfaces to the run-time system and for directlyapplying program transformations where possible.This proposal investigates locality optimizations needed for threefeatures found in advanced scientific applications (3D arrays,irregular accesses, and pointers). It focuses on extendinglocality optimizations to handle cache conflicts between multipledata, deep memory hierarchies (multi-level caches and TLBs), hardwareand software prefetching, nonlinear memory layouts, parallel andcluster architectures, and memory performance tools. Localityoptimizations will be applied to representative programs andexperimentally evaluated on advanced computer systems.The results of this research should help improve the performanceof computationally intensive scientific applications. Becauseof trends in computer architectures, lessons learned are alsolikely to be useful for application domains such as imageprocessing and high-performance databases.