CPA: Towards Designing Complex Systems: Exponential Design/Configuration/Parameter Space Exploration Tools That Are Efficient, Accurate, and Easily Usable

计算系统和应用程序正朝着更复杂的操作模式、更高的抽象层次和更大的执行规模发展。这些趋势挑战了在足够详细的硬件或系统模拟器上对实际工作负载建模的技术状态。计算机系统设计者传统上依靠模拟实验来评估所提出设计的特性（如性能、能耗和热特性）。不幸的是，不断增长的系统复杂性使得许多传统的基于仿真的方法效率低下或过时。底线是，计算机设计和工作量已经变得足够复杂，对它们进行完全详细的建模是完全难以处理的。研究人员和实践者缺乏必要的工具来理解不同硬件组件和软件需求之间的许多相互作用，因此无法保证大型设计空间的粗略建模足以识别任何给定系统中代表最重要权衡的设计点。提出的研究通过建立准确，自信和预测模型的自动化方法来解决指数设计空间和配置空间探索的问题。对设计空间中的采样点进行仿真，并利用仿真结果教导模型描述设计/配置参数之间关系的函数。该模型对空间中其他点的预测结果非常准确，可以通过查询来预测参数变化的影响，并且与模拟相比速度非常快，可以有效地发现不同区域参数之间的权衡。研究议程包括调查减少开发准确预测模型所需的样本点数量的方法，将该方法应用于新的硬件设计空间，以及使用预测建模来共同配置软件工作负载及其运行的硬件。

Computing systems and applications are evolving towards more sophisticated modes of operation, higher levels of abstraction, and larger scales of execution. These trends challenge the state of technology for modeling realistic workloads on sufficiently detailed hardware or system simulators. Computer system designers traditionally rely on simulation experiments to evaluate properties (such as performance, energy consumption, and thermal characteristics) of proposed designs. Unfortunately, growing system complexity renders many traditional simulation-based approaches woefully inefficient or obsolete. The bottom line is that computer designs and workloads have grown sufficiently complex that modeling them fully in detail is utterly intractable.Researchers and practitioners lack the necessary tools to understand the many interactions of diverse hardware components and software requirements, and thus there is no assurance that coarse-scale modeling of large design spaces is sufficient to identify design points representing the most important tradeoffs for any given system.The proposed research attacks the problem of exponential design space and configuration space exploration via an automated approach that builds accurate, confident, and predictive models. Sample points in the design space are simulated, and the results are used to teach the models the function describing relationships among design/configuration parameters. The models produce highly accurate predicted results for other points in the space, can be queried to predict impacts of parameter changes, and are very fast compared to simulation, enabling efficient discovery of tradeoffs among parameters in different regions. The research agenda includes investigating ways to reduce the number of sample points required to develop accurate predictive models, applying the approach to new hardware design spaces, and using predictive modeling to jointly configure software workloads together with the hardware on which they run.