Evaluation of a Hypervisor Performance in a Distributed Embedded System

In modern industrial systems cloud computing plays an important role. Using this technology services are given to customers more efficiently with respect to cost and performance. The main idea of using cloud is that a platform is partitioned among several end users providing a complete isolation among the services. Therefore, the resources are used more effectively instead of assigning a complete platform for only one end user. In order to provide such a partitioning among the services several techniques are already being used. One of the prominent techniques is virtualization. Virtualization can be done with the use of a hypervisor, which is a software that allows running of multiple operating systems on one platform. This thesis aims at measuring the performance of a hypervisor in a distributed embedded system. The metrics for measuring the performance are delays, jitters and throughput that are influenced by different architectures and settings. The thesis also aims at finding out if it is possible to predict the delay, jitter and throughput depending on the number of virtual machines, number of switches and amount of network load. Finally, the thesis investigates whether different settings on the virtual machines influence the performance. For these purposes, a network consisting of two hypervisors and one or two network switches is setup. On each hypervisor several virtual machines are installed. Different tools for measurement of network performance, such as Iperf and Jperf, are installed on the virtual machines. The results show that network load is the main factor influencing the delay, jitter and throughput in the network. The number of switches influence to some degree due to the processing delay. The number of virtual machines has no or very low influence on the network performance. Finally, the results show that alternating configurations on the virtual machines has no observable differences in delay, jitter or throughput, albeit with the limited changes in the settings for this experiment