SysOptic: A Fine-Grained Monitoring System for Virtual Machines Based on PMU

Abstract

Modern cloud datacenters indicate the frequent ex-istence of complex failure manifestation. Failures have becomethe norm, and not the exception. This is a key challengesince assumptions that underpin designing reliable systems aremonitoring systems status and detecting anomaly at runtime. Performance Monitoring Unit on CPU (PMU) can obtainfine-grained monitoring data by adopting interrupt samplingmethod based on hardware events. However, profilers in virtualmachines fail to receive PMU relevant information directlydue to the limited capacity of PMU virtualization. In thispaper, we present a fine-grained monitoring system SysOpticbased on PMU virtualization. First, we propose a method toexpose PMU information PMU and ensure the visibility ofsuch information at virtual machine level. Second, to maximizethe PMU reusability, SysOptic supports the PMU sharing andsimultaneous monitoring among multiple virtual machines. Furthermore, we also describe how to synchronize interruptson physical machines to virtual machines by using injectinginterrupts. Experimental results show that with the aid ofSysOptic, profiler tools in virtual machines manage to perceivethe existence of PMU and collect the monitoring data. Theadditional overhead incurred by SysOptic is at most 9.8%.