CAREER: Rethinking HPC Resilience in the Exascale Era

弹性是高性能计算（HPC）中关键的百亿亿级研究挑战之一。由于高错误率，百亿亿次超级计算机在计算方面几乎没有进展，或者可能由于故障而产生不正确的结果，从而使百亿亿次的性能毫无用处。挑战在于如何在不增加性能开销、功耗和底层硬件复杂性的情况下，在百亿亿次上实现完整的HPC弹性。为此，本研究项目设计并开发了百亿亿次时代hpresilience的低成本硬件/软件协同技术。本项目涉及四个研究目标：(1)低成本的cpu软错误弹性；智能编译器-体系结构交互可以验证错误的缺失并执行细粒度恢复，从而消除SDC。(2)面向编译器的商用gpu软错误弹性；它可以从GPU寄存器文件中删除耗电的纠错码（ECC）逻辑，而不会影响其弹性。(3)轻量级非易失性存储器（NVM）持久性；它可以减轻传统重量级hpc检查点的开销，并支持没有不可撤销操作的应用程序的全系统持久性。(4)在保证程序正确性的前提下，采用低成本的时序误差弹性，实现能源效率最大化。由此产生的人工制品和技术有望通过解决建立可靠的高性能计算系统的挑战，为国家的竞争力做出贡献。该研究成果影响了从嵌入式系统到高性能计算云等需要正确计算结果的可靠计算系统的广泛学科。因此，使用所提出的技术将使当前和新兴应用程序的执行更加可靠，从而直接影响我们的生活方式。该研究项目将产生三种类型的数据：(1)算法和模型，(2)软件原型，(3)测试基础设施，包括模拟器和评估基准及其痕迹，(4)教育材料。我们所有的软件工具都将是开源的，可供公众、实验室和工业界使用。

Resilience is one of the key exascale research challenges in high-performancecomputing (HPC). Due to much high error rates, exascale supercomputers couldmake little progress in computations, or might generate incorrect results due tofailures, rendering the exascale performance useless. Thechallenge is how to achieve a complete HPC resilience at exascale in a way thatdoes not increase the performance overhead, the power consumption, and thecomplexity of underlying hardware. To this end, this research project designsand develops low-cost hardware/software cooperative techniques for HPCresilience in the exascale era. This project involves four research goals: (1) low-cost soft error resiliencefor CPUs; intelligent compiler-architecture interaction can validate the lack oferrors and performs fine-grained recovery, thus eliminating SDC. (2)compiler-directed soft error resilience for commodity GPUs; it can remove thepower-hungry error-correcting code (ECC) logic from the GPU register fileswithout compromising their resilience. (3) lightweight nonvolatile memory (NVM)persistence; it can mitigate the overhead of traditional heavyweight HPCcheckpointing and support whole-system persistence for applications withoutirrevocable operations. (4) low-cost timing error resilience for aggressivevoltage scaling to maximize the energy-efficiency with program correctnessguarantee.The resulting artifacts and technologies are expected to contribute to thenation's competitiveness by addressing the challenge of building reliable HPCsystems. The research outcome impacts a broad range of any disciplines thatneed correct computation results thus requiring reliable computing systemscovering from embedded systems to HPC cloud. Consequently, use of the proposedtechniques will make the execution of current and emerging applications muchmore reliable, and therefore directly affect our way of life.There will be three types of data generated from this research project: (1)algorithms and models, (2) software prototype, (3) testing infrastructureincluding simulators and evaluation benchmarks and their traces, (4) educationalmaterials. All of our software tools will be open source and made available tothe public, laboratories and industry.