We propose ACCT, an end-to-end learning-based congestion control mechanism for Internet flows that aims to achieve an optimal sending rate through collaboration between applications and the network elements along the network path. Specifically, an ACCT sender can exploit explicit feedback from the network about its ability to satisfy the application's desired sending rate. Unlike existing schemes, ACCT does not blindly adjust its sending rate to either of these signals; instead, it considers them advisory because the bottleneck may be at routers that do not support ACCT signaling. To detect such bottlenecks and strictly control the end-to-end delay for latency-sensitive applications, ACCT actively measures the end-to-end queuing delay and regulates its sending rate to keep queuing delay within an acceptable margin. Finally, ACCT heavily relies on a Reinforcement Learning (RL) agent to adjust the aggressiveness of its sending rate. We propose a novel approach to formulate the RL agent, exploiting multiple reward functions simultaneously to train the agent. We implemented the ACCT congestion control module in NS-3, which communicates with the RL agent via a technology-agnostic protocol. We evaluated performance across various network scenarios in both wired and wireless networks. Across all experiments, ACCT significantly outperforms commonly used TCP variants such as CUBIC; it provides network fairness when competing with other ACCT flows, and finally, ACCT flows coexist well when they carry heterogeneous traffic.

We propose XRC, an explicit rate control algorithm that overcomes the poor performance of commonly used TCP variants in cellular networks. XRC exploits explicit feedback from the radio access network that is aware of the physical, network and transport layer information of all UEs as well as resource distribution policies for users with different traffic characteristics. XRC co-exists fairly with other XRC and non-XRC flows at the wireless and non-wireless bottlenecks while it strictly controls queuing delay within a small threshold. We implement XRC in NS-3 and examine its performance across a range of network loads and dynamics. When competing with CUBIC at a wireless bottleneck, XRC achieves a Jain’s fairness index of 99.7% while providing a 3x lower median queuing delay compared to when CUBIC competes with CUBIC in the same setup.

In a packet network, the routes taken by traffic can be determined according to predefined objectives. Assuming that the network conditions remain static and the defined objectives do not change, mathematical tools such as linear programming could be used to solve this routing problem. However, networks can be dynamic or the routing requirements may change. In that context, Reinforcement Learning (RL), which can learn to adapt in dynamic conditions and offers flexibility of behavior through the reward function, presents as a suitable tool to find good routing strategies. In this work, we train an RL agent, which we call R2L, to address the routing problem. The policy function used in R2L is a neural network and we use an evolution strategy algorithm to determine its weights and biases. We tested R2L in two different scenarios: static and dynamic networks conditions. In the first, we used a 16-node network and experimented with different reward functions, observing that R2L was able to adapt its routing behavior accordingly. Finally, in the second experiment, we used a 5-node network topology where a given link's transmission rate changed during the simulation. In this scenario, we observed that R2L was able to deliver a competitive performance, compared to heuristic benchmarks, with changing network conditions.

Today almost half the world's population does not have Internet access. This is particularly the case in rural and undeserved regions where providing Internet access infrastructure is challenging and expensive. To this end, we present demonstration of WhiteHaul [5], a low-cost hybrid cross-layer aggregation system for TV White Space (TVWS) based backhaul. WhiteHaul features a custom-designed frequency conversion substrate that efficiently handles multiple noncontiguous chunks of TVWS spectrum using multiple low-cost COTS 802.11n/ac cards but with a single antenna. At the software layer, WhiteHaul uses MPTCP as a link-level tunnel abstraction to efficiently aggregate multiple chunks of the TVWS spectrum via a novel uncoupled, cross-layer congestion control algorithm. This demo illustrates the unique features of the WhiteHaul system based on a prototype implementation employing a modified version of MPTCP Linux Kernel and a custom-designed conversion substrate. Using this prototype, we highlight the performance of the WhiteHaul system under various configurations and network conditions.

We address the challenge of backhaul connectivity for rural and developing regions, which is essential for universal fixed/mobile Internet access. To this end, we propose to exploit the TV white space (TVWS) spectrum for its attractive properties: low cost, abundance in under-served regions and favorable propagation characteristics. Specifically, we propose a system called WhiteHaul for the efficient aggregation of the TVWS spectrum tailored for the backhaul use case. At the core of WhiteHaul are two key innovations: (i) a TVWS conversion substrate that can efficiently handle multiple non-contiguous chunks of TVWS spectrum using multiple low cost 802.11n/ac cards but with a single antenna; (ii) novel use of MPTCP as a link-level tunnel abstraction and its use for efficiently aggregating multiple chunks of the TVWS spectrum via a novel uncoupled, cross-layer congestion control algorithm. Through extensive evaluations using a prototype implementation of WhiteHaul, we show that: (a) WhiteHaul can aggregate almost the whole of TV band with 3 interfaces and achieve nearly 600Mbps TCP throughput; (b) the WhiteHaul MPTCP congestion control algorithm provides an order of magnitude improvement over state of the art algorithms for typical TVWS backhaul links. We also present additional measurement and simulation based results to evaluate other aspects of the WhiteHaul design.

MPTCP and its ECN-capable variants such as XMP and DCM have recently been introduced to effectively exploit the path diversity of modern data center networks (DCNs). Although these multipath schemes improve overall network throughput compared to single-path schemes due to their fast, host-based, load balancing ability, they failed to address the following two problems: TCP incast and last hop unfairness. Firstly, these mechanisms cause frequent TCP incast collapses when used for workloads with a many-to-one communication pattern, commonly found in DCNs. Secondly, the last hop unfairness problem severely violates network fairness as single-path flows achieve 2-5 times less throughput than multipath flows. To effectively tackle these problems, we propose the Adaptive MultiPath (AMP) congestion control mechanism that quickly detects the onset of these problems and transforms its multipath flow into a single-path flow. Once these problems disappear, AMP safely reverses this transformation and continues data transmission via multiple paths. Our evaluation results under a diverse set of scenarios in a large-scale fat-tree topology demonstrate that AMP is robust to the TCP incast problem and improves network fairness between multipath and single-path flows significantly with no performance loss.

Modern data centres provide large aggregate network capacity and multiple paths among servers. Traffic is very diverse; most of the data is produced by long, bandwidth hungry flows but the large majority of flows, which commonly come with strict deadlines regarding their completion time, are short. It has been shown that TCP is not efficient for any of these types of traffic in modern data centres. More recent protocols such MultiPath TCP (MPTCP) are very efficient for long flows, but are ill-suited for short flows. In this paper, we present Maximum MultiPath TCP (MMPTCP), a novel transport protocol which, compared to TCP and MPTCP, reduces short flows' completion times, while providing excellent goodput to long flows. To do so, MMPTCP runs in two phases; initially, it randomly scatters packets in the network under a single congestion window exploiting all available paths. This is beneficial to latency-sensitive flows. After a specific amount of data is sent, MMPTCP switches to a regular MultiPath TCP mode. MMPTCP is incrementally deployable in existing data centres as it does not require any modifications outside the transport layer and behaves well when competing with legacy TCP and MPTCP flows. Our extensive experimental evaluation in simulated FatTree topologies shows that all design objectives for MMPTCP are met.

In this paper, we introduce MMPTCP, a novel transport protocol which aims at unifying the way data is transported in data centres. MMPTCP runs in two phases; initially, it randomly scatters packets in the network under a single congestion window exploiting all available paths. This is beneficial to latency-sensitive flows. During the second phase, MMPTCP runs in Multi-Path TCP (MPTCP) mode, which has been shown to be very efficient for long flows. Initial evaluation shows that our approach significantly improves short flow completion times while providing high throughput for long flows and high overall network utilisation.