SANDeRS: Smart, Adaptive Compilation for Dark Silicon

我们生活在一个多核时代：计算处理器不再以时钟速度为卖点，而是以核心数量为标志。处理器的能量和功率密度的基本限制将很快把我们进一步推向一个暗硅时代，在这个时代，任何时候只有一小部分芯片可以供电。在这种情况下，在一个芯片上放置更多相同的处理核心（即同质性）没有任何优势。这迫使计算机架构师引入围绕不同处理器构建的异构多核系统，这些处理器具有不同的能量和性能特征，并且每个处理器都专门用于某类应用程序。计算机架构师现在希望软件能够找到方法来释放异构多核的潜力。然而，软件开发人员正在努力应对这种急剧增加的复杂性；而当前的编译器工具，其作用是使软件能够有效地利用底层硬件，根本不足以完成这项任务。为由相同核组成的同质多核构建优化编译器已经是一项艰巨的任务，即使只是以性能为目标（即使程序更快）。通常需要几代编译器才能开始有效地利用处理器的潜力，这时出现了一个新的处理器，进程又开始了。设计高效的编译器启发式来优化异构多核上的能量（即减少能量消耗）和性能将是一项更加艰巨的任务，特别是考虑到不同核和相互连接的微妙交互。即使成功地实现了这一目标，当迁移到新发布的处理器时，编译器设计的任务也必须重新开始。这种永无止境的追赶游戏不可避免地延迟了上市时间，这意味着我们很少能在硬件的生命周期内充分利用它。如果找不到解决方案，我们将面临软件停滞，无法提供可扩展的计算性能——这是过去50年来极大地改变了我们社会的驱动力。我们需要的是一种能够适应未来硬件体系结构变化的方法，并在硬件世代之间提供可伸缩的性能。这个项目恰恰提供了这一点。它将通过结合计算机科学的两个不同领域：并行编译器设计和机器学习来实现这一目标，以开发一种新的能源和性能优化范例。我们的关键见解是，最佳优化策略可以从类似的软件/硬件设置中学习；学习到的知识可以在没有人类参与的情况下不断更新。这个项目将提供这样一个智能的、自适应的编译系统。我们将使用机器学习来获取工作负载、应用程序和底层硬件的知识，测试新的编译策略，学习如何针对每个特定的计算环境优化每个单独的程序，并随着时间的推移不断改进优化启发式。随着应用环境知识的增长，我们的系统将使程序更快、更节能；例如，软件的反应速度会更快，手机的电池寿命会更长。它将缩短软件产品的上市时间，并随着硬件的进步提供可扩展的性能。如果成功，这样的工作计划将有助于解决暗硅迫在眉睫的软件危机，这将有利于学术界和英国工业界，以及世界各地的系统软件研究人员和开发人员。

We live in an era of multi-cores: computing processors are no longer marketed by their clock speeds, they are marked by the number of cores. The fundamental limits of energy and power density of processors will soon push us further into an age of dark-silicon where only a small portion of the chip can be powered at any time. In such a setting, putting more of the same processing cores on a chip (i.e. homogeneity) gives no advantage. This has forced computer architects to introduce heterogeneous many-core systems built around distinct processors -- which have different energy and performance characteristics and each is specialised for a certain class of applications. Computer architects now hope that software will find ways to unlock the potential of heterogeneous many-cores. Software developers, however, are struggling to cope with this dramatic increase in complexity; and the current compiler tools, whose role is to enable software makes effective use of the underlying hardware, are simply inadequate to the task.It is already a daunting task to build optimising compilers for homogeneous multi-cores consisting of identical cores, even just targeting performance (i.e. to make programs faster). It typically takes several generations of a compiler to start to effectively exploit the processor's potential, by which time a new processor appears and the process starts again. It will be a fundamentally more difficult task to design efficient compiler heuristics for optimising energy (i.e. to reduce energy consumption) and performance on heterogeneous many-cores, especially given the subtle interactions of different cores and inter-connections. Even if successfully achieved, the task of compiler design must likely to be started again when moving to a new released processor. This never ending game of catch-up inevitably delays time to market, meaning that we rarely fully exploit the hardware in its lifetime. If no solution is found, we will be faced with software stagnation and will be unable to offer scalable computing performance -- a driving force that has dramatically changed our society over the past 50 years.What is needed is an approach that evolves and adapts to the future hardware architectural change and delivers scalable performance over hardware generations. This project offers precisely that. It will achieve this by bringing together two distinct areas of computer science: parallel compiler design and machine learning to develop a new paradigm for energy and performance optimisation. Our key insight is that the best optimisation strategies can be learned from similar software/hardware settings; and the learnt knowledge can be constantly refreshed without human involvement. This project will deliver such a smart, adaptive compilation system. We will use machine learning to acquire knowledge of workloads, applications and the underlying hardware, testing new compilation strategies, learning how each individual program should be optimised for each specific computing environment, and constantly improving the optimisation heuristics over time. As knowledge of the application environment grows, our system will make programs faster and more energy efficient; for example, software will respond quicker and the battery life will last longer on mobile phones. It will reduce time to market for software products and deliver scalable performance as hardware advances. If successful, such as programme of work will help to the looming software crisis of dark silicon, which will be of benefit to academics and UK industry, and system software researchers and developers worldwide.