Monitoring the Computational Resource Utilization of VNFs In Telecommunication Cloud

The telecommunication industry is accelerating towards a Cloud-native All-IT infrastructure with flat architecture, to meet the requirements in the era of the fifth generation (5G) of mobile communication systems. Various hardware-based network functions are to be virtualized, operated by software, and distributed into Cloud resources as Virtual Network Functions (VNF). However, this transformation brings additional technical challenges. Performance unpredictability of VNFs can be considered as one of the most critical among all, and it is mainly due to the shared resources pooling feature of Cloud Computing. The virtualization layer, in between the Virtual Machines (VM) and host machine, tries to implement some fairness in scheduling the shared resources but there might be delays in availability of resources, which will degrade the performance of any waiting VMs. Continuous monitoring of the resources utilized by the VMs is thus essential in ensuring their individual performances and the performance of any constituted VNF(s). Monitoring the CPU resource utilization can be considered as most important since the other resources are not directly impacted by the virtualization layer. A VNF’s total CPU utilization is the sum of the total CPU utilized by each VM that constitutes it. For every CPU utilized by a VM, the virtualization layer needs to support and schedule that usage and by doing so, utilizes some CPU itself which needs to be accounted to the VM. Since the virtualization layer resides in the host machine, monitoring and accounting for the CPU utilized by the virtualization layer on behalf of any VM would require privileged access to the host machine which is usually not granted by most Cloud service providers; making it impossible to calculate any VM’s total CPU utilization. This thesis’s study presents a solution to determine a VNF’s total CPU utilization without the need for any privileged access to the host machine. Applications running inside a VNF have specific purposes and their behavior is likely to show some pattern; and thus, CPU utilization pattern exist. The solution focuses on modeling any such CPU utilization pattern using machine-learning. Various regression-based analyses is performed to determine the relationship between set of resource utilization metrics gathered from the VNF and the total CPU utilized by that VNF as seen on the host machine. Based on the outcome of the best analysis, a total-CPU-utilized prediction model is devised, and a real-time performance evaluation of the model is also carried out to confirm its accuracy