Hypervisor-Induced Negative Interference in Virtualized Multi-core Platforms: The P4080 Case

Abstract

Modern real-time (RT) applications heavily depend on multi-core processors to meet the increasing computing demands within strict time-constraints. To reduce the cost of overall hardware, most of these applications run along with general purpose workloads on the same hardware platform which may create data contention in shared resources and incur delay for the time critical RT applications. Virtualization of the hardware platform under the supervision of a Virtual Machine Monitor (VMM) or hypervisor allows the isolation of shared resources among concurrent applications from different domains. However, the actual partitioning of hardware resources among the hosted operating systems as well as the hypervisor play a significant role in the performance of the RT applications. In this paper we report results of an experimental investigation for this purpose on a particular multi-core platform, the P4080 developed by Freescale Semiconductor Inc. Using the actual (not simulated) P4080 hardware platform and the Topaz hypervisor, we examined the effect that the partition of the last-level cache (LLC) among a Real-Time Operating System (RTOS), a General Purpose Operating System (GPOS) and the hypervisor has on the performance of the RT tasks. We have identified inverted bell-shaped curves that seem to govern the optimal allocation of resources for P4080.