Large data transfers can result in significant congestion and performance degradation for interactive end-user applications such as web browsing and streaming. While there are existing TCP congestion control algorithms for delivery of large volume data (e.g., LEDBAT, TCP-LP), our results show that these protocols are not effective in cellular networks due to variability in radio channel conditions and the use of cellular schedulers in base stations. We propose Legilimens, an agile TCP variant for cellular downlink transfers, which not only retains desirable properties of existing approaches, but also exploits the properties of the cellular schedulers to estimate load and capacity and addresses the challenges in cellular networks. As a result, Legilimens is able to deliver traffic using only the spare capacity on the downlink. We conduct extensive evaluations of Legilimens in multiple settings-in a large cellular network for real-world performance, on the PhantomNet emulator for controlled experiments, and ns-3 simulator for scaled experiments-all of which demonstrate that Legilimens is superior to existing protocols in transferring large volumes of data without interfering with regular user traffic. Compared to existing low-priority protocols, Legilimens improves the throughput of background flows by 2x on average (up to 5x) without degrading the performance of foreground flows across all the three testbeds.

Though Remote Direct Memory Access (RDMA) promises to reduce datacenter network latencies significantly compared to TCP (e.g., 10x), end-to-end congestion control in the presence of incasts is a challenge. Targeting the full generality of the congestion problem, previous schemes rely on slow, iterative convergence to the appropriate sending rates (e.g., TIMELY takes 50 RTTs). Several papers have shown that even in oversubscribed datacenter networks most congestion occurs at the receiver. Accordingly, we propose a divide-and-specialize approach, called Dart, which isolates the common case of receiver congestion and further subdivides the remaining in-network congestion into the simpler spatially-localized and the harder spatially-dispersed cases. For receiver congestion, we propose direct apportioning of sending rates (DASR) in which a receiver for n senders directs each sender to cut its rate by a factor of n, converging in only one RTT. For the spatially-localized case, Dart provides fast (under one RTT) response by adding novel switch hardware for in-order flow deflection (IOFD) because RDMA disallows packet reordering on which previous load balancing schemes rely. For the uncommon spatially-dispersed case, Dart falls back to DCQCN. Small-scale testbed measurements and at-scale simulations, respectively, show that Dart achieves 60% (2.5x) and 79% (4.8x) lower 99t'-percentile latency, and similar and 58% higher throughput than InfiniBand, and TIMELY and DCQCN.

Datacenter applications frequently cause incast congestion, which degrades short flows' flow completion times and long flows' throughput. Existing congestion control schemes (e.g., DCTCP) do not explicitly detect and isolate incast. Instead, they rely on existing Explicit Congestion Notification (ECN) to react to general congestion. They, therefore, lose performance due to slow, cautious, and inaccurate reaction to incast. We propose a novel algorithm that detects incasts and notifies senders using a new Explicit Incast Notification EIN). We show that our incast detection is fast and accurate. Next, we present our congestion control scheme, called Pulser , which isolates incasts using EIN. Unlike DCTCP, which gradually adjusts sending rate, Pulser drastically backs off during incast and rapidly restores sending rate once incast ends (i.e., like a pulse). Our real experiments and ns-3 simulations show that Pulser outperforms prior schemes, DCTCP and ICTCP, in both flow completion times and throughput.

A large variety of traffic - time-sensitive “foreground” traffic (e.g., web browsing) and time-insensitive “background” traffic (e.g., software updates) - compete for the scarce cellular bandwidth, especially on the downlink. While there is limited in-network support for traffic prioritization, existing endto-end, “low priority transport protocols” exhibit sub-optimal performance in cellular networks. We propose Sneaker, which yields to time-sensitive foreground traffic during periods of congestion and enables time-insensitive background traffic to efficiently utilize any spare capacity. Sneaker achieves the desired goal by randomly dropping packets coming into the base station, based on traffic type and network conditions. Our key contribution is the derivation of the optimal dropping rate and a practical dropping rate, which performs close to optimal. Further, Sneaker co-exists and performs well with existing cellular schedulers and transport protocols.