Harmonizing latency-critical and non latency-critical services while maintaining QoS and energy efficiency
Jørgen Johannessen · NORA - Norwegian Open Research Archives · 2018
With the rapid increase in popularity of cloud computing and all its related technology, challenges start to emerge regarding the energy efficiency or lack thereof in server parks of all shapes and sizes.It is only in the later years of the existence of data centers providers have realized that power consumption, carbon emission and general power costs are becoming to great to ignore, and that energy efficiency measures must be taken to be able to survive the ever increasing demand of computing power.This thesis will investigate how such energy efficiency problems may be handled by exploring multiple ways of managing the power versus performance balance of CPUs by implementing techniques such as dynamic frequency scaling, CPU pinning and CPU hotplugging.Additionally, this project will propose a feedback control loop which will be implemented on a server that will control a virtualized environment with the purpose of achieving fast and precise elasticity with vertical resource scaling to meet certain QoS requirements for our services while at the same time optimizing the energy efficiency of our system.The services themselves are a pair of latency-critical and non latency-critical services which shall co-exist in an environment where resources are constantly being tuned up and down based on the required performance at each iteration of the loop.iii