On the suitability of black-box performance monitoring for SLA-driven cloud provisioning scenarios

Abstract

In recent years, cloud computing has become an increasingly important software delivery paradigm, mainly for reasons of increased scalability. The scalability benefits are accomplished by the capability of autonomously and elastically scaling up or down so that customer preferences (SLAs) can be accommodated. For this, performance monitoring is a prerequisite. Distinction is made between white-box and black-box monitoring techniques: the former involves collecting information about the monitored component by looking at its internals, while the latter only involves observing the components interfaces. In practice, cloud provisioning is commonly based on white-box monitoring. These techniques are costly to develop, since technologies (and providers) offer their own white-box inspection APIs and are costly to integrate (e.g., in a multi-cloud setup involving different providers). In addition it is not always possible to apply this performance monitoring technique when dealing with third-party components or services. In this paper, we investigate whether typical SLA-driven cloud provisioning scenarios can be supported when relying exclusively on black-box performance monitoring techniques. We perform an experiment in which we apply both white-box (e.g., CPU usage, load, etc.) and black-box instrumentation (e.g., latency of operations, amount of failed operations, etc.) in a realistic case study, and we discover clear correlations between some of the obtained white-box and black-box measurements. As such, we show that black-box performance monitoring techniques can be used to support such provisioning scenarios.