A measurement study on virtualization overhead for applications of industrial automation systems

In a remote building management system (BMS) that manages facilities of buildings or factories, there is an application called a crawler that collects statuses of facilities via a network. By running a crawler on a Virtual Machine (VM) for each building owner, operators of the remote BMS can limit influence of a failure of one crawler and can reduce the number of physical machines in the remote BMS. However, performance of the crawler running on a VM could be degraded owing to overhead of virtualization, inappropriate resource allocation, and interference among VMs coexisting on the same physical machine. In addition, the crawler needs to meet its performance requirements with high probability, 99.999%, and therefore it is important to clarify the characteristics of the performance degradation. In this paper, we evaluate performance of the crawler run on a VM with various resource allocation patterns and multiple VMs. From this evaluation, we find out that 1) CPU usage measurement with millisecond granularity is important for appropriate CPU allocation and 2) CPU contention among VMs is one of the main factors of the performance degradation.