Packet Value Based PIE Queue Management Using Programmable Data Planes : Enhancing Active Queue Management Using Packet Value-Based Scheduling in P4

Sammanfattning

Efficient network resource management is critical for maintaining low latency and ensuring Quality of Service (QoS) across diverse traffic types. Traditional Active Queue Management (AQM) algorithms, such as Random Early Detection (RED) and Proportional Integral Controller Enhanced (PIE), fail to account for the inherent value of packets, leading to inefficient prioritization of latency sensitive traffic. To address this, per Packet Value (PV) based queue management has been proposed, allowing packets to be prioritized based on service type, contractual agreements, and network policies. However, integrating such an approach into programmable data planes presents significant challenges, particularly in Programming Protocol independent Packet Processors (P4), which lacks support for floating point operations, recursion, and loops. This thesis investigates the feasibility of implementing the Packet Value Proportional Integral Controller Enhanced (PVPIE) algorithm an extension of PIE that incorporates per PV within a P4 programmable data plane. The implementation overcomes P4’s limitations by employing register approximations for statistical calculations and an alternative method for computing congestion thresholds. The proposed solution was evaluated in a network emulation environment, analyzing key performance metrics such as queue latency, throughput, and fairness. Results demonstrate that P4 Packet Value Proportional Integral Controller Enhanced (P4PVPIE) effectively maintains target queueing delays while prioritizing high value packets under congestion. Compared to PIE and First-In, First-Out (FIFO) queuing, P4PVPIE achieves improved latency control and resource allocation, making it a viable solution for future programmable network infrastructures. However, trade offs in fairness were observed, as lower value traffic experiences increased packet loss. These findings highlight the potential of value based queue management in optimizing QoS while underscoring the challenges of implementing such mechanisms within constrained programmable environments.