Understand Your Chains and Keep Your Deadlines: Introducing Time-constrained Profiling for NFV

Fully-automated resource dimensioning is one of the key requirements for agile, DevOps-enabled network function virtualization (NFV) scenarios in which new service versions are continuously delivered and deployed to production. To enable and support these dimensioning processes, a priori knowledge about the performance behavior of the deployed service function chains (SFC) is collected using profiling solutions that automatically generate so-called SFC performance profiles. A challenge in those profiling processes is the huge configuration space of typical SFCs that need to be explored to collect enough information to accurately describe the performance behavior of the profiled SFC. In this paper, we introduce the concept of time-constrained profiling (T-CP) which profiles only a small subset of all possible SFC configurations and uses machine learning techniques to predict performance values for the remaining configurations to generate a full SFC performance profile within a given time limit. Using our novel, open-source T-CP prototype, we analyze the accuracy of the generated profiles using different selection algorithms to find good configuration subsets. We base parts of this analysis on real-world SFC performance measurements, which we make publicly available as open dataset.