Quantifying Architectural Requirements of Contemporary Extreme-Scale Scientific Applications

摘要: As detailed in recent reports, HPC architectures will continue to change over the next decade an effort improve energy efficiency, reliability, and performance. At this time of significant disruption, it is critically important understand specific application requirements, so that these architectural changes can include features satisfy requirements contemporary extreme-scale scientific applications. To address need, we have developed a methodology supported by toolkit allows us investigate computation, memory, communication behaviors applications at varying levels resolution. Using methodology, performed broad-based, characterization 12 scalable benchmarks. Our analysis reveals numerous sometimes contradict conventional wisdom about For example, results reveal only one our executes more floating-point instructions than other types instructions. In another found topologies are very regular, even for that, first glance, should be highly irregular. These observations emphasize necessity measurement-driven real applications, help prioritize included future architectures.