Performance analysis of general P4 forwarding devices with controller feedback: Single- and multi-data plane cases

Abstract

Software-Defined Networking (SDN) lays the foundation for the operation of future networking applications. The separation of the control plane from the programmable data plane increases the flexibility in network operation, and hence, improves the overall performance. One of the most used languages for describing the packet behavior in the data plane is P4. It allows both protocol and hardware independent programming. With the expanding deployment of P4 programmable devices, it is of utmost importance to understand their achievable performance and limitations in order to design a network and provide Quality of Service (QoS) guarantees in terms of different metrics of interest to users communicating in the network. One of the most important figure of merits is the mean sojourn time of a packet in a P4 device. While previous works already modeled the sojourn time in P4 devices with controller feedback, those models were rather oversimplified and could not capture the real system behavior for general cases, resulting this way in a potentially high inaccuracy in performance prediction. To bridge this gap, in this paper, we consider the system behavior of P4 devices for the general case, i.e., under general input parameter distributions. To that end, we model the system behavior with a queueing network with feedback. First, we do this for a single data plane, and then we extend the analysis to the case when there are multiple data planes sending occasional packets to the same controller. Due to the fact that it is impossible to provide closed-form solutions in the general case, we consider different approximation approaches for the mean sojourn time and show which one is better for a given scenario. We validate our results against extensive realistic simulations, capturing different behaviors in the data and control planes. Results show that the most accurate approximation in most cases is the one in which queueing networks are decoupled and considered as independent queues despite the fact that there are considerable dependencies involved. The level of discrepancy with the best approximating approach in the worst case for a single data plane does not exceed 18.2% for service times distributions with a coefficient of variation not greater than 1, whereas when dealing with multiple data planes, the discrepancy is usually higher, but with the best-approximating approach in each case rarely exceeds 14%.