Energy-efficient Resource Allocation Through Energy Fungibility In Serverless

The massive energy consumption of data centers has become an increasingly significant pain point for cloud providers.However,existing researches only focus on the power constraints of a server to ensure the quality of service.They do not directly design the resource allocation system to reduce server energy consumption.At the same time,serverless has become the computing paradigm of the new generation of cloud services due to its extremely high development efficiency,rapid elastic scaling,and actual usage billing.Although the serverless scenario makes it more challenging to reduce energy consumption,we find that there are still opportunities to reduce the energy consumption of serverless workloads,thereby improving the data center’s energy efficiency.In particular,we describe the detailed energy consumption in the serverless scenario,analyze the energy information to find the part with high energy consumption,and propose a resource allocation scheme to improve energy efficiency.Firstly,we split the cycle of serverless workloads into three stages: cold start,runtime,and idle.More importantly,we observe distinct energy consumption patterns in different stages.Based on the observation,we coordinate the energy used in each stage to reduce overall energy consumption.Secondly,we discover that energy consumption is fungible,e.g.,different combinations of multi-dimensional resources could guarantee the performance of functions.However,the different resource combinations cause entirely different energy consumption.Exploiting energy fungibility,we can find the optimal resource configuration that can guarantee performance and reduce the energy of the function.By balancing the energy consumption and performance of the function,we can reduce the energy consumption without violating the function performance constraints.We propose RAEF,a function-level resource allocation technology for serverless computing scenarios,which can proactively adjust the resource allocation of functions to reduce function energy consumption with ensuring function performances.The experimental results show that RAEF can reduce the energy consumption of the function by up to 21% compared with the optimal frequency adjustment controller.