COKE - Consistency Kernel - Software-Controlled Consistency and Coherency for Many-Core Architectures

现代处理器中不断增加的核心数量越来越多地使传统的、硬件实现的缓存一致性受到质疑。同时，处理器内部高速网络提供了将一致性保持方法从硬件转移到操作和运行时系统的机会。因此，COKE项目第一阶段的中心目标是研究和实现灵活动态共享模型，以及在缓存不一致的多核处理器环境中保持一致性的相关机制。在第二阶段，将继续第一阶段的操作并扩展到分区的全局地址空间（PGAS）。除了内存，该项目还将能源作为额外的可管理资源。多核处理器的能量感知操作所追求的方法在软件设计时就已经生效，以便及早发现能量开销并考虑替代实现。其基础是对程序进行静态分析，以便预测最佳、平均和最坏情况下能源消耗的区域数据。在选定的（编译器生成的）移交点，将估计的消耗提交给操作系统控制单元，以便在运行时主动响应程序下一个执行阶段所期望的负载形状。进程的这种本地特性支持系统中的短期调度。除此之外，对于长期调度，操作系统了解程序加载时能耗的全局特征。在最终测量结果的基础上，软件实现的PGAS系统存储区域一致性协议将被划分为能效等级。通过对软件控制、操作系统支持的全局逻辑地址空间和非顺序应用程序的逻辑共享内存的现有知识进行操作，该项目遇到了这样一个基本假设，即由此创建的决策范围能够在操作系统级别动态优化执行，并得到应用程序的提示和调节螺钉的辅助。挑战在于逻辑共享内存的能效、延迟和吞吐量与依赖于数据和时间的应用程序级访问模式、底层硬件结构和动态操作状态之间的特定关系。同样受到挑战的是（内部开发的）新系统和（国外开发的）遗留系统之间在工程概念和技术的适用性、可移植性和泛化方面的共同关系。

Increasing core numbers in modern processors more and more put classical, hardware-implemented cache coherence into question. At the same time, processor-internal high-speed networks provide the opportunity to shift methods for consistency preservation from hardware to operating and run-time systems. Central objective in the first phase of the COKE project therefore was the investigation and implementation of models for flexible dynamic sharing and the associated mechanisms for consistency preservation in the context of cache-incoherent many-core processors. In the second phase, the operations of the first phase shall be continued and extended to partitioned global address spaces (PGAS).Besides memory, the project foregrounds energy as additional managed resource. The pursued approach for energy-aware operation of many-core processors takes effect already at design time of software to uncover energy overhead early and contemplate for alternate implementations. Basis for this is the static analysis of programs in order to predict region-wise data for best-, average-, and worst-case energy consumption. At selected (compiler-generated) hand-over points, the estimated consumption is committed to an operating-system control unit to proactively respond at run-time on the load shape expected for the next execution phase of a program. This local characteristic of a process supports short-term scheduling in the system. In addition to this, for long-term scheduling, the operating system learns about the global characteristic of energy consumption at load time of the program. On the basis of the finally achieved measuring results, the software-implemented consistency protocols for the memory regions of a PGAS system will be graded into energy-performance classes.By acting on existent knowledge on software-controlled, operating-system supported global logical address spaces and logical shared memory for non-sequential applications the project encounters the basic assumption that the thus created scope of decision enables dynamic optimisation of execution at operating-system level, assisted by hints and regulating screws by the application. Challenged is the specific relation between energy efficiency, latency, and throughput of the logical shared memory and the data- and time-dependent application-level access patterns, the structure of the underlying hardware, and dynamic operational state. Also challenged is the common relation between the (in-house developed) new system and the (foreign-developed) legacy system as to applicability, portability, and generalisation of the engineered concepts and techniques.