Abstract

In this paper, we analyze the problem of overloads caused by physical CPU contention in cloud infrastructures, from the perspective of time-critical applications (such as Virtual Network Functions) running at guest level. We show that guest-level overload control solutions to counteract traffic spikes (e.g., traffic throttling) are counterproductive against overloads caused by CPU contention. We then propose a general guest-level solution to protect applications from overloads also in the case of CPU contention. We reproduced the phenomena on a IP Multimedia Subsystem (IMS) testbed based on OpenStack on top of KVM. The results show that the approach can dynamically adapt the service throughput to the actual system capacity in both cases of traffic spikes and CPU contention, by guaranteeing at the same time the IMS latency requirements. • Resource contention is a common cause of overload in virtualized infrastructures. • CPU contention causes side effects on the QoS of time-critical applications. • Guest OS CPU metrics are often misinterpreted when dealing with CPU contention. • NFV Load throttling solutions can misbehave in the case of physical CPU contention. • Avoiding idle CPU time preemption helps to protect systems from contention effects.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call