CPA-CPL: Exploring and Exploiting Heterogeneous Cache Sharing in Chip Multiprocessors Systems for Locality Optimization and Proactive Cache Management

越来越多的功耗、散热和设计复杂性问题导致了处理器技术向多核多处理器的转变。随着这种转变，内存层次结构的共享变得更加深入、异构和复杂，从而导致缓存争用、冲突增加以及协同共享。如果不了解这种变化的含义，当前的多核系统就会遭受相当大的性能下降、性能隔离不良和公平性保证不足的问题。随着处理器级并行度的快速增加，这些问题的紧迫性也在增加。先前的研究，主要是在架构和操作系统领域，依赖于简单的启发式来估计合并程序的缓存需求；不准确性和开销限制了它们的可扩展性和有效性。本工作通过构建运行进程的预测行为模型，开发缓存共享感知程序转换和循环调度，并将编程系统的程序级知识与运行时系统的主动资源管理相结合，从编译器方面独特地解决了这些挑战。具体来说，这项工作提出了包容性重用签名来表征包容性局部性——在共享缓存上运行程序的内存行为，以及线程间亲和力模型来捕获并行线程之间的数据局部性。它解决了包容性地域性测量、预测和开发所面临的挑战。该分析为编译器和运行时系统的共享缓存优化提供了新的机会。该PI开发了一系列程序转换，如线程间内存重组和缓存共享感知循环调度，以提高线程间空间局部性，改善冲突、争用和虚假共享。对于运行时系统，这项工作发明了主动缓存管理，它根据预测的包容性主动分区缓存或调度进程，克服了当前响应方案在可扩展性，准确性和有效性方面的限制。

The increasing problems of power, heat dissipation, and designcomplexity have caused a shift in processor technology to favormulticore multiprocessors. Along with that shift, the sharing ofmemory hierarchy becomes deeper, heterogeneous and more complex,causing cache contention, increased conflicts, and also, synergysharing. Without understanding the implications of this change,current multicore systems suffer from considerable performancedegradation, poor performance isolation and inferior fairnessguarantees. The urgency of these issues increases as the degree ofprocessor-level parallelism increments rapidly.Prior studies, mostly in areas of architecture and operating systems,rely on simple heuristics to estimate cache requirement of corunningprograms; the inaccuracy and overhead limits their scalability andeffectiveness. This work tackles these challenges uniquely from thecompiler aspect by constructing predictive behavior models forcorunning processes, developing cache-sharing-aware programtransformations and loop scheduling, and combining the program-levelknowledge of programming systems with the proactive resourcemanagement by runtime systems. Specifically, this work proposesinclusive reuse signatures to characterize inclusive locality---thememory behavior of corunning programs on shared caches, andinter-thread affinity models to capture data locality among parallelthreads. It tackles the challenges facing the measurement, predictionand exploitation of inclusive locality. The analysis opens newopportunities for shared-cache optimizations by both compilers andruntime systems. The PI develops a series of program transformations,such as inter-thread memory reorganization and cache-sharing awareloop scheduling, to increase inter-thread spacial locality andameliorate conflicts, contention and false sharing. For runtimesystems, this work invents proactive cache management which partitionscaches or schedules processes according to predicted inclusivelocality proactively, overcoming the limitations of current reactiveschemes on scalability, accuracy and effectiveness.