Energy-efficient I/O for supercomputing

超级计算基础设施已达到工业规模。 我们现代经济所依赖的视频、搜索和云计算服务都是由仓库大小的设施提供的，这些设施容纳了数千台独立的服务器。这些设施的能源消耗在经济和环境方面都是巨大的。 不幸的是，在设计这些分布式超级计算机时，能源效率并不是首要考虑的问题。因此，它们在大多数时间内以其峰值能效的15-65%运行。 该项目将通过针对其中的互连来提高分布式超级计算环境的能源效率。 如果这些互连两端的数字输入/输出（I/O）电路没有进展，它们在整体服务器功率中的份额将在未来十年增加到50%以上。 我们将开发I/O集成电路和子系统，这些集成电路和子系统能够在空闲时接近零能耗，亚纳秒唤醒时间，并可扩展以满足未来的需求。 并行超级计算基础设施的大规模工业规模使这项研究在经济和环境方面产生了巨大的影响。 以单个高端计算服务器中的中央处理器为例，到2020年，它将需要1 TB/s的聚合I/O带宽。 使用今天的技术，仅I/O就将在计算机的5年寿命中消耗大约800美元的电力。 这个研究项目可以从一台计算机上节省750美元的能源-超过CPU的总成本。 对于有环保意识的人来说，这意味着减少了7500公斤的二氧化碳排放量。 每个数据中心都有数千台这样的服务器，全球范围内的数字是惊人的。计算服务器每年消耗的总能量为220 TWh，其中约10%可归因于I/O。 因此，即使是将这些装置的I/O能源效率提高1%的研究，也可以节省相当于20，000个家庭平均用电量的费用。 我们的研究承诺改善远远超过1%。

Supercomputing infrastructure has reached an industrial scale. The video, search, and cloud computing services upon which our modern economy relies are delivered from warehouse-sized facilities housing thousands of individual servers. The energy consumption of these facilities is enormous, both in economic and environmental terms. Unfortunately, energy-efficiency has not been a primary concern in the design of these distributed supercomputers. Consequently, they operate at 15-65% of their peak energy efficiency most of the time. This project will improve the energy-efficiency of distributed supercomputing environments by targeting the interconnections within them. Without progress on the digital input/output (I/O) circuits at either end of these interconnects, their share of overall server power will increase to over 50% in the next decade. We will develop I/O integrated circuits and subsystems that are capable of near-zero energy consumption when idle, sub-nanosecond wakeup times, and scalable to meet the demands of the future. The massive industrial scale of parallel supercomputing infrastructure makes the impact of this research dramatic, both economically and environmentally. Consider the central processor in a single high-end compute server, which by 2020 will demand an aggregate I/O bandwidth of 1TB/s. Using today's technology, the I/O alone will consume roughly $800 in electricity over the computer's 5-year lifetime. This research project can offer a $750 energy savings from a single computer - more than the total CPU cost. For the environmentally minded, this represents a savings of 7500 kg in CO2 emissions. With thousands of such servers in each data center, the worldwide numbers are staggering. The total energy per year consumed by compute servers is 220 TWh, roughly 10% of which is attributable to I/O. Hence, even research that improves I/O energy efficiency by only 1% in these installations yields a savings equivalent to the average electricity consumption of 20,000 homes. Our research promises improvements far exceeding 1%.