An innovative approach to performance metrics calculus in cloud computing environments: a guest-to-host oriented perspective

In virtualized systems, the task of profiling and resource monitoring is not straight-forward. Many datacenters perform CPU overcommittment using hypervisors, running multiple virtual machines on a single computer where the total number of virtual CPUs exceeds the total number of physical CPUs available. From a customer point of view, it could be indeed interesting to know if the purchased service levels are effectively respected by the cloud provider. The innovative approach to performance profiling described in this work is based on the use of virtual performance counters, only recently made available by some hypervisors to their virtual machines, to implement guest-wide profiling. Although it isn't possible for the virtual machine to access Virtual Machine Monitor, with this method it is able to gather interesting informations to deduce the state of resource overcommittment of the virtualization host where it is executed. Tests have been carried out inside the compute nodes of FIWARE Genoa Node, an instance of a widely distributed federated community cloud, based on OpenStack and KVM. AgiLab-DITEN, the laboratory I belonged to and where I conducted my studies, together with TnT-Lab–DITEN and CNIT-GE-Unit designed, installed and configured the whole Genoa Node, that was hosted on DITEN-UniGE equipment rooms. All the software measuring instruments, operating systems and programs used in this research are publicly available and free, and can be easily installed in a micro instance of virtual machine, rapidly deployable also in public clouds.