CAREER: Compiler-Assisted Data Adaptation

数据管理已经成为计算的核心功能，不仅对性能有很大的影响，而且对程序的能耗、程序安全性、容错性等都有很大影响。虽然许多程序使用大量数据，但它们并不总是访问所有数据。因此，数据自适应可以通过精确识别和压缩给定计算点所需的数据来提高缓存效率。这种细粒度、动态化的数据管理方式可以显著缓解通信瓶颈，加强对数据的控制，研究基于编译器的细粒度数据适配方法。它开发了三种编译器辅助技术。首先，程序监控分析选定的数据；然后，局部性分析确定它们的组局部性；最后，数据转换改变它们的布局，同时使用运行时数据映射保持程序的正确性。新系统改进了传统机器上数据密集型应用程序的高速缓存优化，以及嵌入式设备上通用程序的内存管理。在研究的同时，研究人员开发了数据分析工具，以支持计算机组织、编译和编程课程中记忆问题的教学。研究人员还开发了一门高级课程，重点是所有系统级别的内存优化技术。

Data management has become a core function of computing, stronglyinfluencing not only performance but also energy consumption, programsecurity, fault tolerance, and other program properties. While manyprograms use a large amount of data, they do not access all data atall times. Hence, data adaptation can improve cache efficiency byidentifying and compacting precisely those data required at a givenpoint of computation. Such fine-grained and dynamic data managementmay significantly alleviate the bottleneck on communication andstrengthen the control of data.This research studies fine-grained data adaptation through acompiler-based approach. It develops three compiler-assistedtechniques. First, program monitoring analyzes selected data; then,locality analysis identifies their group locality; and finally, datatransformation changes their layout while maintaining programcorrectness using run-time data maps. The new system improves cacheoptimization for data-intensive applications on conventional machinesand memory management for general-purpose programs on embeddeddevices. In parallel to research, the investigator developsdata-analysis tools to support teaching of memory issues in computerorganization, compiler, and programming courses. The investigatoralso develops an advanced course that focuses on memory optimizationtechniques at all system levels.