INTaaS: Provisioning In-band Network Telemetry as a service via online learning

Abstract

Provisioning In-band Network Telemetry as a service for INT-based and INT follow-up applications in an online manner suffers multiple challenges, including control decisions of different INT path algorithms, resources provisioning coupled over time, and the unforeseeable INT query workloads. To overcome these challenges, this study formulates an online non-linear time-varying integer programming problem that maximizes the overall quality of service through algorithm selection and INT query workloads distribution. Specifically, an online learning algorithm is proposed to make the fractional decisions and it uses a primal–dual mechanism by simultaneously minimizing a convex problem and maximizing a concave problem, based on the previously observed inputs. Furthermore, we design a randomized rounding strategy to convert these fractional decisions to integers with an expectation guarantee. We implement our system prototype based on INTCollector upon real devices, i.e., Barefoot Wedge100BF and Inspur Rack. Our rigorous theoretical analysis shows that our INTaaS achieves sub-linear growth on dynamic regret for the optimal loss and incurs sub-linear growth on dynamic fit for the long-term constraint violations. Finally, extensive evaluations on real-world data indicate that INTaaS exhibits up-lift performance up to 40% over other state-of-the-art algorithms.