SHF: Small: Collaborative Research: Experimental-based Research on Effective Models of Parallel Application Execution Time, Power, and Resilience

并行系统的规模和复杂性的不断增长给并行应用带来了巨大的挑战。其中一个挑战是并行应用程序的执行时间、功能和弹性的集成和平衡。MuMMI_R项目旨在促进对各种应用程序系统配置的功率、执行时间和弹性之间相互依赖关系的科学理解。更广泛的影响包括对本科生和研究生的培训，以及参与reu、CREU和DREU等项目，以增加来自代表性不足群体的学生参与该项目。MuMMI_R研究旨在开发有效的技术，用于量化执行时间、功率和弹性之间的复杂权衡，并为用户定义的指标提供调优机制。为了实现这一目标，研究集中在三个相互关联的研究重点：(1)实验研究，在不同并行架构下，对一套具有代表性的应用程序在不同弹性策略下进行广泛的实验；(2)应用级协同建模，建立分析模型和基于彩色Petri网的仿真，量化执行时间、功耗和弹性之间的相关性和权衡；(3)基于模型的分析，检验弹性、执行时间和弹性之间的权衡。以及针对不同的应用程序系统配置的能力，以及针对当前和未来系统上用户定义的目标度量来调优应用程序实现。由此产生的框架MuMMI_R将对应用程序-系统交互提供有价值的见解，并有助于设计高效的并行应用程序（关于执行时间、功率需求和弹性）、运行时系统和计算机体系结构。

The increasing scale and complexity of parallel systems present enormous challenges to parallel applications. One such challenge is the integration and balancing of execution time, power, and resilience for parallel applications. The MuMMI_R project seeks to advance the scientific understanding of the interdependence among power, execution time, and resilience for various application-system configurations. The broader impacts include training of undergraduate and graduate students and the participation in programs such as REUs, CREU, and DREU to increase the participation of students from underrepresented groups in the project.The MuMMI_R research aims to develop effective techniques for quantifying the complicated tradeoffs among execution time, power, and resilience, and to provide a tuning mechanism for user-defined metrics. Toward this goal, the research focuses on three interrelated research thrusts: (1) experimental research to conduct extensive experiments of a suite of representative application under different resilience strategies on various parallel architectures, (2) application-level co-modeling to develop analytical models and colored Petri net based simulation for quantifying the correlations and tradeoffs between execution time, power, and resilience, and (3) model-based analysis to examine the tradeoffs among resilience, execution time, and power for different application-system configurations, and to tune application implementations for a user-defined target metric on current and future systems. The resulting framework, MuMMI_R, will provide valuable insights into application-system interactions and aid in the design of efficient parallel applications (with respect to execution time, power requirements, and resilience), runtime systems, and computer architectures.