Climate change due to increasing carbon emissions by human activities has been identified as one of the most critical threat to Earth. Carbon neutralization, as a key approach to reverse climate change, has triggered the development of new regulations to enforce the economic activities toward low carbon solutions. Computing networks that enable users to process computation-intensive tasks contribute huge amount of carbon emissions due to rising energy consumption. To analyze the achievable reduction of carbon emissions by a scheduling policy, we first propose a novel virtual queueing network model that captures communication and computing procedures in networks. To adapt to highly variable and unpredictable nature of renewable energy utilized by computing networks (i.e., carbon intensity of grid varies by time and location), we propose a novel carbon-intensity based scheduling policy that dynamically schedules computation tasks over clouds via the drift-plus-penalty methodology in Lyapunov optimization. Our numerical analysis using real-world data shows that the proposed policy achieves 54% reduction on the cumulative carbon emissions for AI model training tasks compared to the queue-length based policy.
翻译:由于人类活动导致的碳排放量增加而导致的气候变化被认为是对地球的最严重威胁之一。碳中和作为扭转气候变化的关键方法,已经引发制定新的规章,以强制经济活动,实现低碳解决方案。使用户能够处理计算密集型任务的计算机网络由于能源消耗增加而促成大量碳排放。为了通过时间安排政策分析可实现的碳排放减少,我们首先提议一个新的虚拟排队网络模型,以捕捉网络中的通信和计算程序。为了适应计算网络使用的可再生能源的高度可变和不可预测的性质(即电网的碳密度因时间和地点而异),我们提议了一项基于碳密集度的新计划政策,该规划政策动态地安排通过Lyapunov优化的漂移加线方法对云进行计算任务。我们利用现实世界数据进行的数字分析表明,拟议政策与基于排队长度的政策相比,在AI模式培训任务中累积的碳排放量方面实现了54%的减少。