Current Virtual Machine (VM) scheduling services, such as Openstack’s Nova only have awareness of the CPU, RAM, and storage utilization from a hypervisor perspective. While this has proven to be sufficient for traditional cloud applications and their associated workloads, as they are rarely limited by network bandwidth, Virtual Networks Functions (VNFs) by comparison require fairly static and known quantities of CPU, memory and storage. However, properly placing their workloads depends upon a knowledge of hypervisor network resources and external network topology.
This paper will compare and contrast traditional cloud workloads and their placement with VNF workloads and their respective placement. We will discuss why theoretical extensions to cloud management software are needed to improve placement, quality of service and utilization of commodity hardware.