Performance trade-offs of auto scaling schemes for NFV with reliability requirements

Abstract

A major concern for future mobile networks is to meet extremely stringent reliability requirements without incurring very high energy consumption. To this aim, we need auto-scaling techniques to switch on/off servers on demand, ensuring that they are active when required. One key challenge is to keep the right number of active servers while accounting for both the fallibility of servers and their non-zero boot up time: if this number is too small we risk service disruption, but if it is too large resources will be wasted. In this paper, we analyze this challenge by assessing the performance of different strategies to support an energy-efficient highly-reliable service, which range from deploying few blade (i.e., reliable) servers to deploying many nano (i.e., unreliable) servers. Our main take-away message is two-fold: (1) a server farm of nano servers can provide a more energy efficient operation than a farm of blade servers, but (2) it involves a very dynamic operation both in terms of task migrations and (de)activations, which challenges the endurance of the hardware.