Reducing Noisy-Neighbor Impact with a Fuzzy Affinity-Aware Scheduler

Abstract

Overbooking techniques have been proven efficient to increase overall utilization of cloud datacenters. However, overbooking may also degrade applications performance as (at least) some applications need to share physical resources such as CPU or memory. Consequently, interference may increase among the virtual machines that share resources, the so called noisy neighbors effect. We present an affinity-aware scheduler to reduce the impact of such interference. A fuzzy logic engine accounts for the uncertainty in these environments and estimates which CPU cores are currently more suitable for each incoming application. This helps the scheduler make virtual machine to physical resource mapping decisions, also known as vcpu pinning. An experimental evaluation based on a combination of interactive services and batch applications confirms that our affinity-aware fuzzy scheduler reduces the interference among applications, enabling more predictable performance and consequently safer overbooking.