Routers use queues to temporarily hold packets when they can’t process them immediately. Congestion occurs when the arrival rate of packets exceeds the processing capacity, leading to increased queuing delay. Active Queue Management (AQM) strategies aim to reduce congestion and queuing delay by draining packets from queues. Traditionally, AQMs are placed in the Egress pipeline of Programmable Data Plane (PDP) hardware due to the availability of queue delay information there. We argue that this approach wastes the router’s resources because the dropped packet has already consumed the entire pipeline of the device. We propose ingress Random Early Detection (iRED) as a more efficient P4-AQM approach fully implemented in PDP hardware, that introduces the concept of disaggregated AQM, effectively resolving the egress drop problem. iRED also supports the Low Latency, Low Loss, and Scalable Throughput (L4S) framework, saving device pipeline resources by dropping packets in the Ingress block. Our experiments with a Tofino2 programmable switch achieved a rate of 1Tbps and show that iRED can reduce router resource consumption by up to 10x in memory usage, 12x fewer processing cycles, and 8x less power consumption for the same traffic load. It achieves fairness in bandwidth usage for different types of traffic and significantly improves Quality of Service (QoS) in Dynamic Adaptive Streaming over HTTP (DASH) with up to a 2.34x improvement in Frames Per Second (FPS) and a 4.77x increase in video player buffer fill. We believe this is a pioneering work to highlight the egress drop problem and deeply explore its potential damage to network performance.

This demonstration addresses a key open challenge in Open Radio Access Network (O-RAN) deployments: how to intelligently allocate Transport Network (TN) resources to ensure low-latency for mission-critical applications. The demo emulates a Smart Factory scenario where the time-sensitive control traffic of robotic arms competes with industrial camera broadband video streams. We propose an intelligent transport controller that combines network slicing, Adaptive Neuro-Fuzzy Inference System (ANFIS), and Federated Learning (FL) to dynamically prioritize traffic per slice. The architecture uses $\mathbf{P 4}$ switches for local queue monitoring and real-time resource scheduling. The integration with the O-RAN disaggregated stack is based on Open Air Interface (OAI). Experimental results demonstrate valuable load balancing and buffer occupation reduction in the O-RAN midhaul.

In-band network telemetry (INT), enabled by programmable data planes and the appearance of programming protocol-independent languages such as P4, emerged as a viable approach for network monitoring. INT allows the collection of fine-grained network information in real-time, increasing network visibility, at the cost of network overhead. Several lightweight INT approaches have been recently proposed that attempt to alleviate the transmission overhead of INT, while maintaining a high degree of monitoring accuracy. However, their impact on the resources of the respective hardware network devices has been hardly investigated as most of the approaches are evaluated via simulation. In this study, we provide proof of concept implementations of two lightweight INT approaches that have been proposed for path tracing on the Intel Tofino ASIC, identifying the challenges of porting the solution to the selected target. We examine their performance, providing an in-depth analysis of resource consumption.

This study explores the application of Time Series GAN in a Programmable Data Plane (PDP) for enhancing Reinforcement Learning within the context of computer networks, particularly in video applications. We address various challenges, including dataset augmentation, balancing, and extended RL training times in real setups. By leveraging synthetic data generated by TimeGAN, we accelerate experimentation, enhance dataset diversity, and simplify RL model training, ultimately evaluating TimeGAN’s performance against real setups in resource optimization for PDPs using an RL agent. This research contributes by directly comparing GAN usage and real setups, bridging a gap in computer network literature, and highlighting a 99% similarity in Quality of Service achieved by an RL model trained with synthetic data, affirming TimeGAN’s potential as a valuable simulator without compromising RL training efficacy.

Video services account for the largest share of all Internet traffic, demanding a network capable of supporting the requirements of delay-sensitive traffic. Fluctuations in network load can cause high delays in the queues of network routers, which tend to degrade the Quality of Service (QoS) for adaptive video streaming, such as Dynamic Adaptive Streaming over HTTP (DASH). This work is positioned in the scope of active management queues (AQM) to improve the QoS of a DASH service by means of dropping packets. One traditional AQM that adopts a packet drop policy is Random Early Detection (RED), developed to drain the flow in times of congestion and thus reduce queueing delay. We revisited and implemented a P4-based implementation of RED, named iRED (ingress RED), an algorithm capable of dropping packets at the ingress pipeline, an innovation compared to other AQM strategies based on dropping at the egress. iRED was evaluated in two scenarios. First, we compare iRED against state-of-art AQM algorithms employing egress packet dropping in terms of Round-Trip Time (RTT), throughput and their impact on resources usage. Our findings indicate that iRED outperforms existing P4-based approaches by approximately up to 2.5x in RTT and 0.75x in throughput for the given buffer sizes. Next, we compare iRED versus Tail Drop (TD) approach in an emulated programmable Content Delivery Network (CDN) employing DASH. Experiments indicate that the iRED improve the QoS by approximately 0.85x in terms of cached video available in the client’s buffer and 0.9x in Frames Per Second (FPS) played.