Adam Lackorzynski, TU Dresden

Operating systems such as Linux are increasingly used for real-time work due to their combination of versatility and offering of low event latency. When virtualizing such a system, for example, to consolidate multiple different systems into a single host, the required latency constraints must still be fulfilled. Consequently the hypervisor as well as the virtual machine monitor must be designed in a way that the virtual machine does not block unduly long to violate latency constraints because of communication with the virtual machine monitor or other host system services.

Using embedded systems virtualization requires the paravirtualization approach as full system virtualization is not yet available. Typically paravirtualization approaches are coupled more tightly with the host system by using operating system interfaces. For example, the creation and destruction of address spaces must be done by the host kernel and thus the paravirtualized parts of the virtual machine must use hypervisor calls for handling address spaces. However, creating and especially destroying address spaces can be time costly and thus have a negative impact on the event latency of the virtual machine. The host adaption of a virtual machine must therefore never let the virtual machine block unduly long in such hypercalls. Nevertheless address spaces must be managed and the calls must be done. Putting those in a different host thread within the virtual machine turned out to have problems, for example, because the thread needs to have a  lower host priority which in turn leads to unbound resource usage.

The paper will detail on those observations and describe a possible solution that maintains a low event priority in the system.