The Kelvin Living Lab: Towards Net Zero High-Performance Computing

该项目的目的是评估将开尔文-2高性能计算(HPC)的碳排放减少到零或负的备选方案。这将通过提高效率、实施调度和补充技术来实现。HPC耗电量大，并且有相应的冷却要求，这可能会显著增加功耗。因此，瞄准加工效率和废热为脱碳提供了最大的机会。拟议的产出侧重于这些备选方案，力求证明能效有效提高40%，并将由此产生的指导方针提供给其他二线中心。我们的目标是通过多管齐下的方法提高HPC的可持续性：1.分析和部署软件和硬件，以提高基础设施的利用率和能效在过去20年里，各种研究提出了提高HPC系统能效的方法，如并行应用的有效调度、动态并行度控制和动态电压和频率调节，包括PI和De Supinski的研究，例如EP/M01147X/1、EP/L000555/1和EP/M015742/1。然而，由于服务上的未知风险，这些技术尚未从研究过渡到实践。我们的目标是在开尔文-2系统的单独切片上重新评估和调整这些技术。此外，我们还将探索其他节点体系结构，因为不同的节点体系结构可能会为特定应用提供更好的能效。此外，我们还将调查是否可以利用动态电压和频率调节来帮助维持电力公用事业电网的稳定性，因为可再生能源降低了电网应对动态稳定性问题的能力，这限制了它们的采用。基于时间同步仪器(OpenPMU)的专业知识，我们将探索如何实时适应高性能计算基础设施的功耗，以应对电网频率波动。用户可持续参与对于HPC中心来说，如何激励用户可持续地执行工作是一个悬而未决的问题。然而，我们可以告知用户其工作对环境的影响，例如，通过报告能源和碳使用量。缺少的环节是运营策略和调度算法，这些策略和调度算法在考虑环境影响的情况下对作业进行优先排序。这需要一个公平的收费方案，反映执行工作的时间和能源权衡的最佳性(避免因为大型节能工作的规模而惩罚它们)。净零氟氯烃利用可再生能源的运行可以显著减少与数据处理相关的排放。通过将HPC与其他利用废热来捕获或减少其他过程排放的技术相结合，它们可以真正实现净零排放。使用温度数据，我们将研究各种热能脱碳选项，例如评估新型的直接空气碳捕获(DAC)和与热消费者共处一地。我们将通过与其他高性能计算设施和利益相关者共享知识来最大限度地发挥该项目的影响，这将通过与北爱尔兰高性能计算(https://ni-hpc.ac.uk/)，)、eFutures 3.0电子网络以及行业合作伙伴戴尔技术、AlcesFlight和北爱尔兰电力网络的合作而得到支持。

The purpose of this project is to evaluate options which reduce the carbon emissions of the Kelvin-2 high-performance computing (HPC) to zero or negative. This will be achieved by improving efficiency, and implementing scheduling and complementary technologies. HPC is a large electricity consumer and has a commensurate cooling requirement which can significantly increase power consumption. Therefore, targeting the processing efficiency and waste heat offers the greatest opportunity to decarbonise. The proposed outputs focus on these options and seek to demonstrate an effective 40% improvement in energy-efficiency with the resulting guidelines made available to other Tier-2 centres. We aim to improve the sustainability of HPC through a multi-pronged approach:1. Analysing and deploying software and hardware to improve utilisation and energy-efficiency of the infrastructureIn the last two decades, various research has proposed ways to improve energy-efficiency of HPC systems, such as efficient scheduling of parallel applications, dynamic parallelism control and dynamic voltage and frequency scaling, including research by the PI and de Supinski, e.g., EP/M01147X/1, EP/L000555/1 and EP/M015742/1. However, these techniques have not transitioned from research to practise due to unknown risks on the service. Our goal is to re-evaluate and adjust such techniques on an isolated slice of the Kelvin-2 system. Additionally, we will explore alternative node architectures as different node architectures may present better energy-efficiency for specific applications. We will moreover investigate if we can leverage dynamic voltage and frequency scaling to help maintain the electrical utility grid's stability, as renewable energy sources reduce the grid's ability to counter dynamic stability problems, which limits their adoption. Based on expertise in time-synchronised instrumentation (OpenPMU), we will explore adaptation of power consumption of HPC infrastructures in real time to counter electrical grid frequency fluctuations.2. User engagement with sustainabilityIt is an open question for HPC centres how to incentivise users to execute jobs sustainably. We can however inform users on the environmental impact of their jobs, e.g., through reporting energy and carbon usage. The missing links are operational policies and scheduling algorithms that prioritise jobs with environmental impact in mind. This requires a fair charging scheme that reflects on the optimality of the time-vs-energy trade-off of executing jobs (avoiding penalising large energy-efficient jobs because of their size).3. Operations for Net zero HPCUtilising renewable sources can reduce the emissions associated with data processing significantly. By integrating HPC with other technologies which utilise the waste heat to capture or reduce the emissions of other processes they can be truly net zero. Using temperature data we will investigate various options for thermal enabled decarbonisation such as evaluation of novel direct air carbon capture (DAC) and co-location with heat consumers.We will maximise the impact of this project through knowledge sharing with other HPC facilities and stakeholders, which will be supported through collaborations with Northern Ireland HPC (https://ni-hpc.ac.uk/), the eFutures 3.0 network on electronics, and industry partners Dell Technologies, AlcesFlight and Northern Ireland Electricity Networks.