EAGER: Resilient, Energy Efficient HPC System Configuration

高性能计算系统(超级计算机)在规模和复杂性上持续增长。今天的前沿系统包含数万个服务器节点，而拟议的下一代系统可能包含数十万个节点。在这种规模下，当硬件组件可能每隔几分钟或几个小时发生故障时，维护系统运行变得越来越困难。不断增加的系统规模带来了围绕能源可用性和成本的互补挑战，预计系统将消耗10兆瓦或更多的电力。为了使未来的高性能计算系统具有可用性和成本效益，我们必须开发新的设计方法和运行原则，以体现大规模系统的两个重要现实：(A)频繁的硬件组件故障是正常运行的一部分；(B)必须与性能和弹性一样仔细地管理能源消耗和电力成本。作为这项研究的一部分，首席研究员将把商业云计算的新想法应用到HPC系统中，重点放在可靠性和能效上。这包括基于适当规模的硬件构建块的高性能计算系统设计模型，以平衡组件更换和维修的运营成本与过度配置的资本成本，并将能源成本和约束纳入调度系统和资源分配，使研究人员能够看到计算成本。针对关键国家利益的科学、工程和国防问题的超大规模计算系统的部署，目前受到系统可靠性和能源消耗的限制。可靠性和能源管理的新设计和操作方法既可以降低成本，又可以增加访问，使计算机公司能够设计更大的系统，研究机构可以更广泛地部署系统，研究人员可以更好地管理计算资源。

High-performance computing systems (supercomputers) continue to grow in size and complexity. Today's leading edge systems contain tens of thousands of server nodes, and proposed, next-generation systems are likely to contain hundreds of thousands of nodes. At this scale, maintaining system operation when hardware components may fail every few minutes or hours is increasingly difficult. Increasing system sizes bring a complementary challenge surrounding energy availability and costs, with projected systems expected to consume ten or more megawatts of power. For future high-performance computing systems to be useable and cost effective, we must develop new design methodologies and operating principles that embody the two important realities of large-scale systems: frequent hardware component failures are a part of normal operation and (b) energy consumption and power costs must be managed as carefully as performance and resilience. As part of this research, the principal investigator will apply new ideas from commercial cloud computing to HPC systems, focusing on reliability and energy efficiency. This includes models of high-performance computing system design based on right-sizing hardware building blocks to balance operating costs for component replacement and repair against capital costs for over-provisioning, and incorporation of energy costs and constraints into scheduling systems and resource allocations, making computing costs visible to researchers. The deployment of very large-scale computing systems, which target science, engineering and defense problems of critical national interest, is currently limited by both system reliability and energy consumption. New design and operating approaches for reliability and energy management can both reduce costs and increase access, allowing computer companies to design larger systems, research institutions to deploy systems more widely, and researchers to better manage computational resources.