Building Fuel Powered Supercomputing Data Center at Low Cost
Building Fuel Powered Supercomputing Data Center at Low Cost
复制标题
DOI：
10.1145/2751205.2751215
复制
发表时间：
2015-06
期刊：
影响因子：
--
通讯作者：
Yiqing Hua;Chao Li;W. Tang;Li Jiang;Xiaoyao Liang
中科院分区：
文献类型：
--
作者：
Yiqing Hua;Chao Li;W. Tang;Li Jiang;Xiaoyao Liang
以各种经济的清洁燃料为能源的分布式发电系统正在成为超大规模计算系统的有前途的电源。近年来，由于对减少IT碳足迹和服务器能源成本的需求不断增加，这些非传统电源在数据中心设计中的采用越来越多。然而，这种燃料驱动的数据中心的好处往往受到其高初始资本成本（CapEx）的严重影响。这是因为目前大多数试点设计要么依赖昂贵的先进发电机，要么采用具有昂贵备用电源的低性能发电机。在这项研究中，我们利用异构发电，以减少燃料驱动的数据中心的成本。我们发现，不同类型的电源，如果一起使用，可以大大提高成本效益的自我发电，但引入了一个新的设计复杂性层，并提出了一个重要的问题，如何分配计算任务的异构电源。具体而言，由于异构发电机的非理想的输出功率响应速度，服务器可能会招致严重的功率预算不足时，调度大量的工作。我们将这种现象称为电力滞后，这会危及系统的可靠性，并且通过昂贵的电力备用系统来处理是不经济的。为了克服这一障碍，我们提出批处理，一个敏捷的负载调度方案，消除电源滞后在系统/软件级别。除了在不考虑电力系统行为的情况下批量调度计算任务之外，Batch智能地将作业队列分成小集合，并基于功率斜坡率约束和总功率预算约束递增地调度作业。使用现实的HPC数据中心负载跟踪，我们证明了批处理使超级计算机能够在异构电源上平稳运行。我们的设计可帮助数据中心运营商节省80%以上的成本，同时保持所需的工作负载性能。
Distributed power generations that fed with various economical clean fuels are emerging as promising power supplies for extremescale computing systems. Recent years have witnessed a growing adoption of these non-conventional power supplies in data center designs due to the heightening demand for reducing IT carbon footprint and server energy cost. However, the benefits of such a fuel powered data center are often severely compromised by its high initial capital cost (CapEx). This is because most pilot designs today either rely on expensive advanced generators or employ low-performance generators with costly standby power backup. In this study we exploit heterogeneous generation to reduce the cost of data center powered by fuel. We show that different types of power supplies, if used together, can greatly improve the cost-effectiveness of self-generation but introduce a new layer of design complexity and raise an important question of how to dispatch computing tasks on heterogeneous power supplies. Specifically, due to the non-ideal output power response speed of heterogeneous generators, servers may incur serious power budget deficiencies when dispatching large amount of jobs. We refer to this phenomenon as power lagging, which jeopardizes system reliability and are not economical to be handled by costly power backup systems. To overcome this barrier, we propose Batch, an agile load dispatching scheme that eliminate power lagging at the system/software level. Other than dispatch computing tasks in bulk without considering power system behaviors, Batch intelligently splits job queue into small sets and incrementally schedule jobs based on the power ramping rate constraints and total power budget constraints. Using realistic HPC datacenter load traces, we demonstrate that Batch enables supercomputers to smoothly operate on heterogeneous power. Our design helps data center operators save over 80% cost while maintaining the desired workload performance.