The Access Traffic Steering, Switching, and Splitting (ATSSS) technology, currently under standardization by the 3rd Generation Partnership Project (3GPP), is designed to enhance wireless access through the bundling of both 3GPP and non-3GPP access. One implementation option involves tunneling data through the Multipath QUIC (MP-QUIC) protocol. The decision of which access network each packet is transmitted on can greatly impact the performance of a network flow, and may also impact the performance of other flows. While Active Queue Management (AQM) in the network could normally manage such interactions, such techniques are limited in a shared multipath-tunnel context, since all traffic in the tunnel share a single connection. This paper introduces a multi-flow, multipath tunneling framework using MP-QUIC. It employs four distinct scheduling policies for packet scheduling decisions. We investigate how such scheduling policies interact with each other, and how they may be combined in order to achieve trade-offs in terms of latency, interactivity, and throughput. Through an extensive investigation in both emulated and real-world environments, we show that Quality of Service (QoS) can be significantly improved by smart combinations of scheduling policies.

Modern internet users expect seamless, uninterrupted communication when using real-time applications, even when sharing the network connection with devices that produce heavy traffic. This trend has pushed the demand for more sophisticated packet schedulers on routers. However, given the limited resources on these routers, it is increasingly important to reduce the overhead associated with these schedulers. This paper presents XDP Queuing (XDQ), our ongoing work on achieving faster packet scheduling on the Linux operating system, which is popular on networking equipment. Linux provides eXpress Data Path (XDP), a high-performance programmable network data path using the eBPF framework, which allows code to process packets early from the driver. While XDP has found numerous uses in the industry, such as Denial-of-Service attack mitigation, load-balancers, and intrusion prevention systems, it currently has no mechanism for queueing or reordering packets and cannot implement traffic scheduling policies. Our contribution, XDQ, is a programmable packet scheduling extension for the XDP framework. XDQ uses recently proposed schemes for programmable queues. It allows writing packet schedulers using eBPF while benefiting from the XDP fast data path. 

The lack of consideration for application delay requirements in standard loss-based congestion control algorithms (CCAs) has motivated the proposal of several alternative CCAs. As such, Copa is one of the most recent and promising CCAs, and it has attracted attention from both academia andindustry. The delay performance of Copa is governed by amostly static latency-throughput tradeoff parameter, δ. However,a static δ parameter makes it difficult for Copa to achieve consistent delay and throughput over a range of bottleneck bandwidths. In particular, the coexistence of 4G and 5G networks and the wide range of bandwidths experienced in NG-RANs can result in inconsistent CCA performance. To this end, we propose a modification to Copa, Copa-D, that dynamically tunes δ to achieve a consistent delay performance. We evaluate the modification over emulated fixed, 4G, and 5G bottlenecks. The results show that Copa-D achieves consistent delay with minimal impact on throughput in fixed capacity bottlenecks. Copa-D also allows a more intuitive way of specifying the latency-throughput tradeoff and achieves more accurate and predictable delay invariable cellular bottleneck.

Multipath connectivity and aggregation of multiple communication links is actively being researched with the aim to achieve higher throughput and lower latency. In this work we perform an emulation-based evaluation of the relative goodput of MPTCP and TCP in a 5G usage context. A large range of path capacity and delay conditions is explored, for both the primary and secondary paths, with over 2000 different configurations evaluated. Evaluations are performed over eight combinations of MPTCP schedulers and congestion controls. The results show that MPTCP running over two links provide lower goodput than TCP over a single link for the majority of cases. Asymmetry in link conditions is in many cases a major complication for the MPTCP scheduler. To examine the predictability of poor performance, and to obtain further insight on the structure of this phenomena, we perform regression modeling of the relative good put. In addition to the traditional approaches of Linear Regression and Random Forest, we also employ Sym-bolic Regression to obtain mathematical expressions capable of providing insight on the path conditions most contributing to poor MPTCP performance. Such regression expressions can be informative when evaluating different schedulers or link aggregation approaches. 

The deployment of the next generation of cellular networks (5G) is expanding to enable new services and improve the quality of existing ones. Despite the benefits of 5G networks, they also present new challenges for the performance of classical and recently-proposed congestion control algorithms (CCAs), e.g., Bottleneck bandwidth and round-trip propagation time (BBR) and Copa. Despite their successful adoption in the Internet, BBR has been shown to overestimate the bottleneck bandwidth in cellular networks, and Copa has not been independently tested on a similar scale and detail as BBR in cellular networks. In this work, we compare the performance of these fairly recent CCAs as well as the widely deployed CUBIC CCA, and a modification to BBR for cellular networks (RBBR) at 5G rates. The evaluation is performed using the emerging QUIC protocol and uses both emulations and live experiments. Our results show that in 5G networks, CUBIC, BBR, and Copa suffer from significant bufferbloat, longer packet delays, and lower throughput, respectively. We also observe that in cases where the bottleneck is largely in the 5G link, RBBR can offer a significant delay reduction compared to BBR and CUBIC. 

BBR is a promising new congestion control algorithm (CCA) that has been shown to result in significantly lower latency compared to conventional loss-based CCAs. However, in cellular networks, where there is a high variability in the available rate, BBR does not perform as well as expected. In such scenarios, BBR tends to overestimate the available capacity and create queues that cause longer packet delays. In this work, we propose Receiver-driven BBR (RBBR), a modified version of BBR that uses rate estimates made at the receiver side rather than at the sender side. We employ a Kalman filter to make a more accurate estimate of the available bandwidth, and we implement the algorithm in QUIC. An evaluation of the proposed CCA is done through extensive 4G trace-based emulations, real 4G network tests and mmWave trace-based emulations representing a 5G scenario. The results show that RBBR is able to achieve an RTT reduction of up to 80\% with a worst-case throughput loss of about 30\%. The results also show that in real 4G networks, RBBR flows experience a more predictable and consistent RTT than what BBR flows do.

Cellular networks have evolved to support high peak bitrates with low loss rates as observed by the higherlayers. However, applications and services running over cellular networks are now facing other difficult congestion-related challenges, most notably a highly variable link capacity and bufferbloat. To overcome theseissues and improve performance of network traffic in 4G/5G cellular networks, a number of in-network and end-to-end solutions have been proposed. Fairness between interacting congestion control algorithms (CCAs) has played an important role in the type of CCAs considered for research and deployment. The placement of content closer to the user and the allocation of per-user queues in cellular networks has increased the likelihood of a cellular access bottleneck and reduced the extent of flow interaction between multiple users. This has resulted in renewed interest in end-to-end CCAs for cellular networks by opening up room for researchand exploration. In this work, we present end-to-end CCAs that target a high throughput and a low latency over highly variable network links, and classify them according to the way they address the congestion control. The work also discusses the deployability of the algorithms. In addition, we provide insights into possible future research directions, such as coping with a higher degree of variability, interaction of CCAs in as hared bottleneck, and avenues for synergized research, such as CCAs assisted by software defined networking and network function virtualization. We hope that this work will serve as a starting point for systematically navigating through the expanding number of cellular CCAs.

Cellular networks are continuously evolving to allow improved throughput and low latency performance for applications. However, it has been shown that, due to buffer over-provisioning, TCP’s standard loss-based congestion control algorithms (CCAs) can cause long delays in cellular networks. The QUIC transport protocol and the Bottleneck Bandwidth and Round-trip propagation time (BBR) congestion control are both proposed in response to shortcomings observed in TCP and loss-based CCAs. Despite its notable advantages, BBR can experience suboptimal delay performance in cellular networks due to one of its underlying design choices: the maximum bandwidth filter at the sender. In this work, we leverage QUIC’s extensibility to enhance BBR. Instead of using the ACK rate observed at the sender side, we apply a more fitting delivery rate calculated at the receiver. Our 5G-trace-based emulation experiments in CloudLab suggest that our modified QUIC could significantly improve latency without any notable effect on the throughput: In particular, in some of our experiments, we observe up to 39% reduction of the round-trip time (RTT) with a worst case throughput reduction of 2.7%.

Internet traffic is comprised of data flows from various applications with unique traffic characteristics. For many cloud applications, end-to-end latency is a primary factor affecting the perceived user experience. As packet losses cause delays in the communication they impact user experience, making efficient handling of packet losses an important function of transport layer protocols. Multipath TCP (MPTCP) is a modification to TCP that enables simultaneous use of several paths for a TCP flow. MPTCP is known to improve throughput. However, the performance of MPTCP is not optimal when handling certain loss scenarios. Efficient packet loss recovery is thus important to achieve desirable flow completion times for interactive cloud-based applications. In this paper we evaluate the performance of MPTCP in handling tail losses using traffic traces from various cloud-based applications. Tail losses, losses that occur at the end of a flow or traffic burst, are particularly challenging from a latency perspective as they are difficult to detect and recover in a timely manner. Tail losses in TCP are handled by using a tail loss probe (TLP) mechanism which was adapted to MPTCP from TCP. We investigate the performance of TLP in MPTCP, comparing the standard implementation to a recently proposed, less conservative approach. Our experimental results show that a less conservative implementation of TLP performs significantly better than the standard implementation in handling tail losses, reducing the average burst completion time of cloud based applications when tail loss occurs by up to 50% in certain cases.

We present the latency-aware multipath schedulerZQTRTT that takes advantage of the multipath opportunities ininformation-centric networking. The goal of the scheduler is touse the (single) lowest latency path for transaction-oriented flows,and use multiple paths for bulk data flows. A new estimatorcalled zero queue time ratio is used for scheduling over multiplepaths. The objective is to distribute the flow over the paths sothat the zero queue time ratio is equal on the paths, that is,so that each path is ‘pushed’ equally hard by the flow withoutcreating unwanted queueing. We make an initial evaluation usingsimulation that shows that the scheduler meets our objectives.