Inference of virtual network functions' state via analysis of the CPU behavior

The on-going process of softwarization of IT networks promises to reduce the operational and management costs of network infrastructures by replacing hardware middleboxes with equivalent pieces of code executed on general-purpose servers. Alongside the benefits from the operator's perspective, new strategies to provide the network's resources to users are arising. Following the principle of "everything as a service", multiple tenants can access the required resources - typically CPUs, NICs, or RAM - according to a Service-Level Agreement. However, tenants' applications may require a complex and expensive measurement infrastructure to continuously monitor the network function's state. Although the application's specific behavior is unknown (and often opaque to the infrastructure owner), the software nature of (virtual) network functions (VNFs) may be the key to infer the behavior of the high-level functions by accessing low-level information, which is still under the control of the operating system and therefore of the infrastructure owner. As such, in the scenario of software VNFs executed on COTS servers, the underlying CPU's behavior can be used as the sole predictor for the high-level VNF state without explicit in-network measurements: in this paper, we develop a novel methodology to infer high-level characteristics such as throughput or packet loss using CPU data instead of network measurements. Our methodology consists of (i) experimentally analyzing the behavior of a CPU that executes a VNF under different loads, (ii) extracting a correlation between the CPU footprint and the high-level application state, and (iii) use this knowledge to detect the previously mentioned network metrics. Our code and datasets are publicly available.