CSR: Small:Conductor: A Run-Time System for Exascale Computing

在本世纪末达到亿级计算目标的关键问题之一-如果不是关键问题--就是电源问题。艾级系统的目标功率限制是20兆瓦，尽管今天的千兆瓦系统-其性能至少比预期的艾级系统低两个数量级-通常消耗约5兆瓦。单靠硬件改进并不能弥合这一鸿沟。PI正在开发一种名为Conductor的运行时系统来解决电力问题。导体的总体目标是在规定的功率范围内产生接近最佳的应用性能。Conductor通过在节点之间和节点内分配功率来实现这一点。首先，PI正在设计和实现一种新的技术，称为功率调度，它解决了节点间的情况。Conductor的第二部分使用功率门控来解决节点内的情况，它允许以比目前架构中普遍可用的更精细的方式关闭单个组件的电源。最后，PI正在研究允许用户通过注释帮助Conductor在仅有功率调度和功率门控不足以实现高性能的情况下实现高性能的技术。本提案中描述的研究的影响是显著的。Conductor对于在非平凡应用程序上实现艾级性能至关重要，它将有助于推动艾级目标的实现。实现艾级性能是国家的重要优先事项，将影响许多应用领域。PI还将研究思想融入到本科生和研究生的并行和分布式计算课程中。

One of the critical problems---if not the critical problem---inreaching the exascale computing goal by the end of the decade is thepower problem. Exascale systems have a target power constraint of 20megawatts even though today's petascale systems---which haveperformance at least two orders of magnitude below prospectiveexascale systems---generally consume around 5 megawatts. Hardwareimprovements alone will not bridge this gap.The PI is developing a run-time system called Conductor to address thepower issue. The overall goal of Conductor is to produce near-optimalapplication performance under a prescribed power bound. Conductorcarries this out by allocating power both between and within nodes.First, the PI is designing and implementing a new technique calledpower scheduling, which addresses the inter-node case. The secondpart of Conductor addresses the intra-node case with power gating,which allows powering off of individual components in a morefine-grain manner than is generally available in architectures today.Finally, the PI is investigating techniques to allow users to assistConductor, through annotations, in achieving high performance in caseswhere power scheduling and power gating alone are not sufficient.The impact of the research described in this proposal is significant.Conductor is essential to achieving exascale performance on nontrivialapplications, and it will help push towards the exascale goal.Achieving exascale performance is an important national priority andwill impact many application domains. The PI is also integrating theresearch ideas into both undergraduate and graduate parallel anddistributed computing courses.