Simplifying Computer Performance Evaluation using Workload Characterization

计算机系统从硬件和软件的角度都变得越来越复杂。由于在系统构建之前难以评估系统的性能和功率，设计这些系统变得越来越具有挑战性。但是，除非在早期设计空间分析期间能够快速估计性能和功率，否则不可能确定好的设计点并设计正确的系统。在微处理器和计算机系统设计界，仿真是硅前性能分析的实际方法，然而，通常需要几天，几周，有时甚至几个月才能用普通的仿真器来模拟一些设计选择。这项研究将探讨下一代计算系统有效设计的性能评估方法。通过这个项目，将开发一个工作负载提取器，以捕获工作负载的基本属性，并创建微型程序序列，以帮助在硅前设计探索期间评估性能和功耗。具体而言，将努力实现以下目标：（i）创建高效和可管理的基准，（ii）捕捉新兴工作负载的本质以进行功率和性能建模，（iii）开发一种方法来创建用于未来系统和工作负载的性能估计的可扩展基准，（iv）开发用于多核系统的基准测试方法，（v）开发用于理解工作负载的基准相似性度量和聚类技术，以及（vi）开发用于使用基准相似性度量来预测应用程序的性能的方法。所提出的多核系统的可扩展基准和评估方法将有助于设计人员在设计下一代计算机系统。 此外，这项研究将导致培训几个研究生在一个领域，这是至关重要的，以保持我们国家的优势，在计算机系统的设计。

Computer systems are becoming increasingly complex both from hardware and software perspectives. Designing these systems is becoming increasingly challenging due to the difficulty in evaluating performance and power of the systems before the system is built. But unless performance and power can be quickly estimated during early design space analysis, it is impossible to identify good design points and design the right systems. Simulation is the de facto method of presilicon performance analysis in the microprocessor and computer system design community, however, often it takes days, weeks and sometimes months to simulate a few design choices using common simulators. The proposed research involves investigating performance evaluation methodologies for effective design of next generation computing systems.Through this project, a workload distiller will be developed to capture essential properties of workloads and create miniature program sequences to help evaluate performance and power during presilicon design exploration. Specifically, the following objectives will be pursued: (i) creation of efficient and manageable benchmarks, (ii) capture of the essence of emerging workloads for power and performance modeling, (iii) development of a methodology to create scalable benchmarks for performance estimation of futuristic systems and workloads, (iv) development of a benchmarking methodology for multicore systems, (v) development of benchmark similarity metrics and clustering techniques for understanding workloads, and (vi) development of a methodology for predicting performance of applications using benchmark similarity metrics. The proposed scalable benchmarks and the evaluation methodologies for multicore systems will help designers during the design of next generation computer systems. In addition, this research will result in training several graduate students in an area that is critical to maintaining our nation's edge in the design of computer systems.