Large language models (LLMs) are being increasingly deployed in healthcare, influencing diagnostic reasoning and clinical workflows. However, evidence of clinician engagement with these systems, how they prompt, constrain, and verify output, remains scarce, particularly in low- and middle-income countries (LMICs). We conducted a mixed-methods study with physicians in Pakistan: (1) logging their interactions while they solved expert-designed clinical vignettes with optional LLM assistance, and (2) interviewing 12 participants about generative-AI-supported diagnosis. Findings highlight diverse prompting strategies from role assignment to cautious scaffolding, with consistent insistence on human oversight. Interviews reveal pragmatic enthusiasm for LLMs as a “second brain” in resource-constrained settings, tempered by skepticism about reliability, privacy, and patient trust. This study contributes evidence of physician-LLM interaction patterns in an LMIC context, a taxonomy of prompting strategies and oversight mechanisms, and design implications for responsible AI integration in healthcare workflows.

Active learning (AL) techniques reduce labeling costs for training neural machine translation (NMT) models by selecting smaller representative subsets from unlabeled data for annotation. Diversity sampling techniques select heterogeneous instances, while uncertainty sampling methods select instances with the highest model uncertainty. Both approaches have limitations - diversity methods may extract varied but trivial examples, while uncertainty sampling can yield repetitive, uninformative instances. To bridge this gap, we propose Hybrid Uncertainty and Diversity Sampling (HUDS), an AL strategy for domain adaptation in NMT that combines uncertainty and diversity for sentence selection. HUDS computes uncertainty scores for unlabeled sentences and subsequently stratifies them. It then clusters sentence embeddings within each stratum and computes diversity scores by distance to the centroid. A weighted hybrid score that combines uncertainty and diversity is then used to select the top instances for annotation in each AL iteration. Experiments on multi-domain German-English and French-English datasets demonstrate the better performance of HUDS over other strong AL baselines. We analyze the sentence selection with HUDS and show that it prioritizes diverse instances having high model uncertainty for annotation in early AL iterations.

We present the first large-scale analysis of the adoption of third-party serving infrastructures in government digital services. Drawing from data collected across 61 countries spanning every continent and region, capturing over 82% of the world's Internet population, we examine the preferred hosting models for public-facing government sites and associated resources. Leveraging this dataset, we analyze government hosting strategies, cross-border dependencies, and the level of centralization in government web services. Among other findings, we show that governments predominantly rely on third-party infrastructure for data delivery, although this varies significantly, with even neighboring countries showing contrasting patterns. Despite a preference for third-party hosting solutions, most government URLs in our study are served from domestic servers, although again with significant regional variation. Looking at overseas located servers, while the majority are found in North America and Western Europe, we note some interesting bilateral relationships (e.g., with 79% of Mexico's government URLs being served from the US, and 26% of China's government URLs from Japan). This research contributes to understanding the evolving landscape of serving infrastructures in the government sector, and the choices governments make between leveraging third-party solutions and maintaining control over users' access to their services and information.

Popular video streaming platforms attract a large number of global marketers who use the platform to advertise their services. While benefiting platforms and advertisers, users are burdened with the costs of advertisements. Users not only pay for these ads with their invested time and personal information, but also through a substantial amount of data translating into direct financial cost. The financial cost becomes even more pronounced in developing countries, where the cost of mobile broadband can be disproportionately high relative to average income levels. In this paper, we perform the first independent and empirical analysis of the data costs of mobile video ads on YouTube, the most popular video platform, from the users' perspective. To do so, we collect and analyze a data set of over 46,000 YouTube video ads. We find that streaming video ads have multiplelatent andavoidable sources of data wastage, which can lead to excessive data consumption by users. We also conduct an affordability analysis to quantify the overall impact of data wastage and reveal the specific data costs per country associated with these losses. Our findings highlight the need for video platform providers, such as YouTube, to minimize data wastage linked to ads, to make their services more affordable and inclusive.

As a popular choice for video and entertainment streaming, YouTube hosts a large audience, including children, who form a growing proportion of its users. Despite separate "made for kids" labelling and stricter moderation of these videos, inappropriate advertising remains a concern as it threatens the safety of YouTube for young viewers. This paper is the first comparative measurement study that explores how advertisement exposure and content vary across child-oriented videos on YouTube. We do this by conducting a cross-regional advertisement analysis on highly viewed "made for kids" labelled content across a total of ten countries with varying regulation. A second front of comparison is carried out between ad patterns on unlabelled and labelled child-oriented videos. Our analysis reveals that the safety of a child's YouTube experience is shaped significantly by their external environment. There also appears to be lax enforcement of YouTube ad and child protection policies, indicated by the presence of unlabelled child-oriented content with weak ad regulation. We discuss the implications of inappropriate exposure on children and suggest policy and implementation measures to mitigate this threat.

Today's Web remains too expensive for many Internet users, especially in developing regions. Unfortunately, the rising complexity of the Web makes affordability an even bigger concern as it stands to limit users' access to Internet services. We propose a novel framework and a fairness metric for rethinking Web architecture for affordability and inclusion. Our proposed framework systematically adapts Web complexity based on geographic variations in mobile broadband prices and income levels. We conduct a cross-country analysis of 99 countries, showing that our framework can better balance affordability and webpage quality while preserving user privacy. To adapt Web complexity, our framework solves an optimization problem to produce webpages that maximize page quality while reducing the webpage to a given target size.

Public service websites act as official gateways to services provided by governments. Many of these websites are essential for citizens to receive reliable information and online government services. However, the lack of affordability of mobile broadband services in many developing countries and the rising complexity of websites create barriers for citizens in accessing these government websites. This paper presents the first large-scale analysis of the affordability of public service websites in developing countries. We do this by collecting a corpus of 1900 public service websites, including public websites from nine developing countries and for comparison websites from nine developed countries. Our investigation is driven by website complexity analysis as well as evaluation through a recently proposed affordability index. Our analysis reveals that, in general, public service websites in developing countries do not meet the affordability target set by the UN’s Broadband Commission. However, we show that several countries can be brought within or closer to the affordability target by implementing webpage optimizations to reduce page sizes. We also discuss policy interventions that can help make access to public service website more affordable.

The popularity of video streaming on smartphones has led to rising demands for high-quality mobile video streaming. Consequently, we are observing growing support for higher resolution videos (e.g., HD, FHD, QHD) and higher video frame rates (e.g., 48 FPS, 60 FPS). However, supporting high-quality video streaming on smartphones introduces new challenges---besides the available network capacity, the smartphone itself can become a bottleneck due to resource constraints, such as low available memory. In this paper, we conduct an in-depth investigation of memory usage on smartphones and its impacts on mobile video streaming. Our investigation - driven by a combination of a user study, user survey, and experiments on real smartphones - reveals that (i) most smartphones observe memory pressure (i.e., low available memory scenarios), (ii) memory pressure can have a significant impact on mobile video QoE when streaming high-quality videos, e.g., resulting in the mean frame drop rate of 9--100% across smartphones and significantly lower user ratings, and (iii) the drop in mobile video QoE happens primarily due to the way in which video processes interact with kernel-level memory management mechanism, with opportunities for improving mobile video QoE through better adaptation by video clients.

The rapid growth in the number of entry-level smartphones and mobile broadband subscriptions in developing countries has served as a motivation for several projects focused on improving mobile users' quality of experience (QoE). One such initiative is the development of Android Go, a customized operating system designed to run over entry-level smartphones. Today, more than 80% entry-level Android smartphones run Android Go. Despite its growing popularity, its effectiveness in improving the Web QoE remains unclear. This paper presents the first independent empirical analysis of Android Go's causal impact on mobile Web performance. We use a combination of controlled experiments and a set of methodological approaches from the econometrics literature to find unbiased estimates of the average causal effect. Our analysis provides insights that have implications for different stakeholders in the ecosystem of entry-level devices.

With the advancement in natural-sounding speech production models, it is becoming important to develop models that can detect spoofed audios. Synthesized speech models do not explicitly account for all factors affecting speech production, such as the shape, size and structure of a speaker's vocal tract. In this paper, we hypothesize that due to practical limitations of audio corpora (including size, distribution, and balance of variables like gender, age, and accents), there exist certain phonemes that synthesized models are not able to replicate as well as the human articulation system and such phonemes differ in their spectral characteristics from bonafide speech. To discover such phonemes and quantify their effectiveness in distinguishing between spoofed and bonafide speech, we use a deep learning model with self-attention, and analyze the attention weights of the trained model. We use the ASVSpoof2019 dataset for our analysis and find that the attention mechanism picks most on fricatives: /S/,/SH/, nasals: /M/,/N/, vowels: /Y/, and stops: /D/. Furthermore, we obtain 7.54% EER on train and 11.98% on dev data when using only the top-16 most attended phonemes from input audio, better than when any other phoneme classes are used.

Today's Web remains too expensive for many Internet users, especially in developing regions. Unfortunately, the rising complexity of the Web makes affordability an even bigger concern as it stands to limit users' access to Internet services. We propose a novel framework and fairness metric for rethinking Web architecture for affordability and inclusion. Our framework provides systematic guidelines for adapting Web complexity based on geographic variations in mobile broadband prices and income levels. Preliminary evaluation shows the resulting architecture can achieve a better balance between Web quality and affordability while preserving user privacy.

Mobile devices have become the primary mode for Internet access in developing countries. Yet typical data plans and SMS costs can be overwhelming for low income users in these countries. In this paper, we explore the design and usability of a free but extremely low bit rate communication channel to address this challenge. We propose, a data communication channel that uses to transmit messages between phones, thereby sacrificing performance in exchange for low cost. While the data rate of is extremely low (<1 bps), our prototype implementation and small scale user studies explore the feasibility of this idea for different types of messaging scenarios. Our results show that could be a viable option for messaging scenarios that require short, pre-determined responses (e.g., survey questions) while for traditional SMS-style messaging, a suitable user interface and other customizations are likely required to make it a viable option for users.

Web Light is a transcoding service introduced by Google to show lighter and faster webpages to users searching on slow mobile clients. The service detects slow clients (e.g., users on 2G) and tries to convert webpages on the fly into a version optimized for these clients. Web Light claims to significantly reduce page load times, save user data, and substantially increase traffic to such webpages. However, there are several concerns around this service, including, its effectiveness in, preserving relevant content on a page, showing third-party advertisements, improving user performance as well as privacy concerns for users and publishers.

Duplication can be a powerful strategy for overcoming stragglers in cloud services, but is often used conservatively because of the risk of overloading the system. We call for making duplication a first-class concept in cloud systems, and make two contributions in this regard. First, we present duplicate-aware scheduling or DAS, an aggressive duplication policy that duplicates every job, but keeps the system safe by providing suitable support (prioritization and purging) at multiple layers of the cloud system. Second, we present the D-Stage abstraction, which supports DAS and other duplication policies across diverse layers of a cloud system (e.g., network, storage, etc.). The D-Stage abstraction decouples the duplication policy from the mechanism, and facilitates working with legacy layers of a system. Using this abstraction, we evaluate the benefits of DAS for two data parallel applications (HDFS, an in-memory workload generator) and a network function (Snort-based IDS cluster). Our experiments on the public cloud and Emulab show that DAS is safe to use, and the tail latency improvement holds across a wide range of workloads.

Existing flow scheduling schemes for data center networks optimize for a specific workload and performance metric. In this paper, we present 2D, a new scheduling policy that offers robustness across performance metrics and changing workloads - a ground existing scheduling policies are unable to cover. 2D combines basic scheduling building blocks of multiplexing and serialization in a principled way, ensuring tail optimal performance across workloads while also improving the average (and lower percentiles) completion times.

IoT deployments often require communication between devices that employ heterogeneous wireless technologies. Traditionally, expensive gateways are used to relay packets between heterogeneous nodes. Recent cross-technology communication offers a low bandwidth alternative, which is only feasible when communication between such nodes is limited to simple binary commands. In contrast, our work capitalizes on the increasing presence of multi-standard radio chips in mainstream IoT devices, to provide a new perspective on how to enable direct communication between heterogeneous nodes. We design Scylla---a software control layer---that allows multiple wireless stacks to coexist on top of a single radio chip, thereby simultaneously offering multiple communication interfaces. Uniquely, Scylla achieves near stack-native performance and requires no changes to the standards.

We present C-Saw, a system that measures Internet censorship by offering data-driven censorship circumvention to users. The adaptive circumvention capability of C-Saw incentivizes users to opt-in by offering small page load times (PLTs). As users crowdsource, the measurement data gets richer, offering greater insights into censorship mechanisms over a wider region, and in turn leading to even better circumvention capabilities. C-Saw incorporates user consent in its design by measuring only those URLs that a user actually visits. Using a cross-platform implementation of C-Saw, we show that it is effective at collecting and disseminating censorship measurements, selecting circumvention approaches, and optimizing user experience. C-Saw improves the average PLT by up to 48% and 63% over Lantern and Tor, respectively. We demonstrate the feasibility of a large-scale deployment of C-Saw with a pilot study.

Online advertising plays a critical role in enabling the free Web by allowing publishers to monetize their services. However, the rise in internet censorship events globally poses an economic threat to the advertising ecosystem. This paper studies this interplay and presents Advention, a system that provides censorship circumvention while serving relevant ads. Advention leverages the observation that ad systems are usually hosted on domains that are different from the publisher domains and are almost always uncensored. Taking cue from this, Advention fetches ads via the direct, uncensored, channel between users and the ad system. Preliminary results show that Advention not only offers high ad relevance compared to other popular relay-based circumvention tools, it also offers smaller page load times.

To achieve scalable performance, datacenter applications (e.g., search and social networking) are designed to have high fanout. However, such a design leads to frequent fabric congestion (e.g., due to incast, imperfect hashing) even when the utilization is low. Such fabric congestion exhibits spatial (e.g., within a rack and across racks) as well as temporal variations. Unfortunately, current approaches infer congestion by focusing on a localized view leading to non-optimal performance. We propose RecFlow, a receiver-based proactive congestion control scheme that uses OpenFlow and ACK spacing to dynamically track changing bottlenecks and reduces buffer overflows while maintaining fairness and high link utilization. Experimental results show that compared to the state-of-the-art, RecFlow achieves negligible packet loss and high goodput while sharing the link capacity fairly between flows.

Datacenter networks are often structured as multi-rooted trees to provide high bisection bandwidth at low cost. To utilize the available bisection bandwidth, an efficient load balancing algorithm is required. Packet Spraying is known to perform well in symmetric topologies as it provides per-packet load balancing over equal cost paths. However, packet spraying performs poorly in asymmetric topologies. In this paper we ask, “How can we make packet spraying effective in asymmetric topologies while retaining its simplicity?” Towards this end, we propose SAPS, “Symmetric Adaptive Packet Spraying”, an SDN-based scheme that uses packet spraying over symmetric virtual topologies. SAPS is based on the key insight that if we provide each flow with a symmetric view of the network fabric, then packet spraying can produce near-optimal performance. We evaluate SAPS using simulations and testbed experiments. Our results indicate that SAPS performs well for a variety of application workloads and asymmetric network scenarios.

Several data center transport protocols have been proposed in recent years (e.g., DCTCP, PDQ, and pFabric). In this paper, we first identify the underlying strategies used by the existing data center transports, namely, in-network Prioritization (used in pFabric), Arbitration (used in PDQ), and Self-adjusting at Endpoints (PASE) (used in DCTCP). We show that these strategies are complimentary to each other, rather than substitutes, as they have different strengths and can address each other's limitations. Unfortunately, prior data center transports use only one of these strategies. As a result, they either achieve near-optimal performance or deployment friendliness (i.e., require no changes to the data plane) but not both. Based on this insight, we design a data center transport protocol called PASE, which carefully synthesizes these strategies by assigning different transport responsibilities to each strategy. The key advantage of PASE over prior art is that it achieves both near-optimal performance as well as deployment friendliness. PASE does not require any changes in network switches (hardware or software); yet, it achieves comparable, or even better, performance than the state-of-the-art protocols (such as pFabric) that require changes to network elements. Our evaluation results show that the PASE performs well for a wide range of application workloads and network settings.

Wireless LANs (WLANs) carry a diversemix of traffic, ranging from delay-sensitive real-time applications to bulk transfers. Using existing QoS mechanisms in high speed WLANs (e.g., 802.11n/ac) presents a tradeoff between maximizing the performance of real-time applications and achieving high throughput. We propose SlickFi, a service differentiation scheme for high-speed WLANs that addresses this tradeoff by leveraging existing dual radio WiFi access points (APs). SlickFi opportunistically adapts channel width on a per-frame basis based on application demands and communicates traffic information across radios to control aggregate frame sizes for improving channel efficiency. SlickFi can be readily deployed on commodity devices using only driver-level changes at the AP-side. We implemented SlickFi in the ath9k driver and using real testbed experiments, show that it can improve aggregate throughput by up to 1.8x and 2.2x over 802.11n and 802.11e, respectively and the PSNR of 1080p videos by up to 3.4x and 1.7x over 802.11n and 802.11e, respectively.

In this paper, we make a case for a redundancy-aware network stack (RANS) for data centers. In RANS, applications expose information about replicas to the network, which in turn, uses duplicate requests to improve performance of typical applications by enabling them to effectively avoid stragglers. At the heart of RANS is the use of duplicate-aware scheduling, which ensures that duplicate-requests do not overload the system and disturb any primary requests. We highlight the challenges and opportunities present at different layers of RANS, from new interfaces that capture replicas and their semantics, to in-network mechanisms that deal with duplicates. Our preliminary evaluation shows the promise of duplicate-aware scheduling in improving performance of typical data center applications.

TCP incast congestion happens in many-to-one communication workflow patterns that frequently arise in large-scale datacenter applications such as web search, social networks, and cluster-based storage systems. Incast congestion can severely degrade the performance of applications. This paper studies the effectiveness of randomizing the TCP retransmission timeout (RTO) in mitigating the impact of incast. Our design is based on the observation that under incast, retransmitted packets also get synchronized due to the use of similar RTOs by the senders. Using analysis and experimental evaluation, we show that there exists a tradeoff between the randomization interval (from which the RTO values are picked) and the number of senders involved in incast. Motivated by this insight, we propose three algorithms (TDA, MAA, and FSA) for the dynamic adaptation of the randomization interval that rely on (a) successive timeouts, (b) explicit knowledge of the level of multiplexing, and/or (c) the knowledge of flow sizes (i.e., large interval for long flows and a small interval for short flows), respectively. Our results show that these algorithms improve goodput by 1.5x-11x for up to 64 senders and provide greater improvement for larger number of senders. The proposed algorithms can be readily deployed as they do not require any changes in switches or applications.

Existing research on Internet censorship primarily focuses on either measurements or circumvention. Considering these two in isolation often leads to designs with limited capabilities: Circumvention is not driven by measurement data and end users find little incentive to help gather such data. We present the preliminary design and implementation of C-Saw, a platform that offers both. The circumvention capability of C-Saw incentivizes consumers to opt-in. As more and more consumers crowdsource, the monitoring data gets richer. This, in turn, offers greater insights into the censorship mechanisms over a wider region, offering even better circumvention capabilities. C-Saw is a browser-based platform set up as a lightweight client-side proxy. C-Saw adapts its circumvention approach to the particular censorship mechanism deployed by a user's ISP, achieving a better balance of circumvention effectiveness and performance. In addition, it can also leverage the heterogeneity in filtering mechanisms across ISPs to achieve better circumvention performance. We demonstrate this using page load times across various ISPs and locations in a censored region. Unlike previous measurement approaches, C-Saw does not require the knowledge of a target URL to be tested. In fact, as URLs get blocked, their information can be monitored in real time.

We propose eSDN; a practical approach for deploying new datacenter transports without requiring any changes to the switches. eSDN uses light-weight SDN controllers at the end-hosts for querying network state. It obviates the need for statistics collection by a centralized controller especially on short timescales. We show that eSDN can scale well and allow a range of datacenter transports to be realized.

The emergence of cloud data centers has led to the design of customized transport protocols such as Data Center TCP (DCTCP). These protocols improve the performance of cloud applications by explicitly accounting for the unique network and traffic characteristics in data centers. However, such protocols have only been evaluated under greenfield deployment scenarios, where the entire data center is assumed to use the same protocol, which may not always be desirable or feasible. This leads to scenarios where these protocols coexist with TCP and thus share the same network resources. This paper considers such scenarios and presents a comprehensive study of the coexistence of DCTCP and TCP. In particular, we evaluate their bandwidth sharing properties under different active queue management schemes (AQM) including RED, DCTCP AQM, and CHOKe. Our results show that under the DCTCP AQM, DCTCP can starve TCP flows. This problem is mitigated through the use of RED, however, significant unfairness remains. Interestingly, we find that CHOKe exacerbates this unfairness. We show that a modified version of CHOKe considerably improves fairness by more accurately penalizing dominating flows.

A fundamental problem in 802.11 wireless networks is to accurately determine the cause of packet losses. This becomes increasingly important as wireless data rates scale to Gbps, where lack of loss differentiation leads to higher loss in throughput. Recent and upcoming high-speed WLAN standards, such as 802.11n and 802.11ac, use frame aggregation and block acknowledgements for achieving efficient communication. This paper presents BLMon, a framework for loss differentiation, that uses loss patterns within aggregate frames and aggregate frame retries to achieve accurate and low overhead loss differentiation. Towards this end, we carry out a detailed measurement study on a real testbed to ascertain the differences in loss patterns due to noise, collisions, and hidden nodes. We then devise metrics to quantitatively capture these differences. Finally, we design BLMon, which collectively uses these metrics to infer the cause of loss without requiring any out-of-band communication, protocol changes, or customized hardware support. BLMon can be readily deployed on commodity devices using only driver-level changes at the sender-side. We implement BLMon in the ath9k driver and using real testbed experiments, show that it can provide up to 5× improvement in throughput.

Many data center transports have been proposed in recent times (e.g., DCTCP, PDQ, pFabric, etc). Contrary to the common perception that they are competitors (i.e., protocol A vs. protocol B), we claim that the underlying strategies used in these protocols are, in fact, complementary. Based on this insight, we design PASE, a transport framework that synthesizes existing transport strategies, namely, self-adjusting endpoints (used in TCP style protocols), innetwork prioritization (used in pFabric), and arbitration (used in PDQ). PASE is deployment friendly: it does not require any changes to the network fabric; yet, its performance is comparable to, or better than, the state-of-the-art protocols that require changes to network elements (e.g., pFabric). We evaluate PASE using simulations and testbed experiments. Our results show that PASE performs well for a wide range of application workloads and network settings.

Data center operators face extreme challenges in simultaneously providing low latency for short flows, high throughput for long flows, and high burst tolerance. We propose a buffer management strategy that addresses these challenges by isolating short and long flows into separate buffers, sizing these buffers based on flow requirements, and scheduling packets to meet different flow-level objectives. Our design provides new opportunities for performance improvements that complement transport layer optimisations.

Today router buffers are sized according to the well-known Bandwidth-Delay Product (BDP) rule, which uses the average round-trip time (RTT) of flows traversing a router. The BDP rule not only leads to large queueing delays, but also imposes “one (buffer) size fits all” philosophy for flows exhibiting a large variation in RTT. When short and long RTT flows compete at a single buffer, they may adversely affect each other in throughput and delay. We propose SplitBuff, a scheme using which short RTT flows achieve low delay and long RTT flows achieve high throughput, without requiring any protocol modifications. With SplitBuff, a router splits a buffer into multiple buffers of varying sizes and maps flows onto these buffers based on their RTTs. We describe SplitBuff and evaluate its performance using extensive ns-2 simulations to demonstrate its effectiveness.

Today's data centers face extreme challenges in providing low latency for online services such as web search, social networking, and recommendation systems. Achieving low latency is important as it impacts user experience, which in turn impacts operator revenue. However, most current congestion control protocols approximate Processor Sharing (PS), which is known to be sub-optimal for minimizing latency. In this paper, we propose Router Assisted Capacity Sharing (RACS), a data center transport protocol that minimizes flow completion times by approximating the Shortest Remaining Processing Time (SRPT) scheduling policy, which is known to be optimal, in a distributed manner. With RACS, flows are assigned weights which determine their relative priority and thus the rate assigned to them. By changing these weights, RACS can approximate a range of scheduling disciplines. Through extensive ns-2 simulations, we demonstrate that RACS outperforms TCP, DCTCP, and RCP in data center environments. In particular, it improves completion times by up to 95% over TCP, 88% over DCTCP, and 80% over RCP. Our results also show that RACS can outperform deadline-aware transport protocols for typical data center workloads.

For provisioning large-scale online applications such as web search, social networks and advertisement systems, data centers face extreme challenges in providing low latency for short flows (that result from end-user actions) and high throughput for background flows (that are needed to maintain data consistency and structure across massively distributed systems). We propose L2DCT, a practical data center transport protocol that targets a reduction in flow completion times for short flows by approximating the Least Attained Service (LAS) scheduling discipline, without requiring any changes in application software or router hardware, and without adversely affecting the long flows. L2DCT can co-exist with TCP and works by adapting flow rates to the extent of network congestion inferred via Explicit Congestion Notification (ECN) marking, a feature widely supported by the installed router base. Though L2DCT is deadline unaware, our results indicate that, for typical data center traffic patterns and deadlines and over a wide range of traffic load, its deadline miss rate is consistently smaller compared to existing deadline-driven data center transport protocols. L2DCT reduces the mean flow completion time by up to 50% over DCTCP and by up to 95% over TCP. In addition, it reduces the completion for 99th percentile flows by 37% over DCTCP. We present the design and analysis of L2DCT, evaluate its performance, and discuss an implementation built upon standard Linux protocol stack.

To achieve high system efficiency with increasing speeds, recent WiFi standards, such as IEEE 802.11e/n, allow burst transmissions with block acknowledgements, provided the initial packet is successfully received. Consequently, a user can sometimes improve its throughput by sending the initial packet at a lower rate than other users. We model such a system as a game. Our results show that the socially optimal strategy is to send the initial packet at a lower rate than the rest of the burst. Such a strategy results in a better Nash Equilibrium than using the same rate for the entire burst. Moreover, we show that using the rate that maximizes the per-packet throughput, as commonly done, can result in performance that is far from the social optimum.

Many promising congestion control protocols use explicit feedback from the network to achieve high performance. These protocols often use congestion signals whose computation requires an estimate of link capacity. Such estimates are not available in networks where capacity varies over time. This paper studies the impact of inaccurate capacity estimates on the performance of congestion control protocols over variable capacity links. As a case study, we focus on 802.11 WLANs. We show that such estimates can lead to either under-utilization or unfairness and network overload. Using a model, we characterize the available capacity of a node in a 802.11 WLAN and then study a method for capacity estimation. Using simulations, we show that the method leads to high utilization and fairness over shared, multi-access networks.

Many congestion control protocols use explicit feedback from the network to achieve high performance. Most of these either require more bits for feedback than are available in the IP header or incur performance limitations due to inaccurate congestion feedback. There has been recent interest in protocols that obtain high-resolution estimates of congestion by combining the explicit congestion notification (ECN) marks of multiple packets, and using this to guide multiplicative increase, additive increase, multiplicative decrease (MI-AI-MD) window adaptation. This paper studies the potential of such approaches, both analytically and by simulation. The evaluation focuses on a new protocol called Binary Marking Congestion Control (BMCC). It is shown that these schemes can quickly acquire unused capacity, quickly approach a fair rate distribution, and have relatively smooth sending rates, even on high bandwidth-delay product networks. This is achieved while maintaining low average queue length and negligible packet loss. Using extensive simulations, we show that BMCC outperforms XCP, VCP, MLCP, CUBIC, CTCP, SACK, and in some cases RCP, in terms of average flow completion times. Suggestions are also given for the incremental deployment of BMCC.

This paper proposes a framework for congestion control, called binary marking congestion control (BMCC) for high bandwidth-delay product networks. The basic components of BMCC are i) a packet marking scheme for obtaining high resolution congestion estimates using the existing bits available in the IP header for explicit congestion notification (ECN) and ii) a set of load-dependent control laws that use these congestion estimates to achieve efficient and fair bandwidth allocations on high bandwidth-delay product networks, while maintaining a low persistent queue length and negligible packet loss rate. We present analytical models that predict and provide insights into the convergence properties of the protocol. Using extensive packet-level simulations, we assess the efficacy of BMCC and perform comparisons with several proposed schemes. BMCC outperforms VCP, MLCP, XCP, SACK+RED/ECN and in some cases RCP, in terms of average flow completion times for typical Internet flow sizes.

Load factor based congestion control schemes have shown to enhance network performance, in terms of utilization, packet loss and delay. In these schemes, using more accurate representation of network load levels is likely to lead to a more efficient way of communicating congestion information to hosts. Increasing the amount of congestion information, however, may end up adversely affecting the performance of the network. This paper focuses on this trade-off and addresses two important and challenging questions: (i) How many congestion levels should be represented by the feedback signal to provide near-optimal performance? and (ii) What window adjustment policies must be in place to ensure robustness in the face of congestion and achieve efficient and fair bandwidth allocations in high bandwidth-delay product (BDP) networks, while keeping low queues and negligible packet drop rates? Based on theoretical analysis and simulations, our results show that 3-bit feedback is sufficient for achieving near-optimal rate convergence to an efficient bandwidth allocation. While the performance gap between 2-bit and 3-bit schemes is large, gains follow the law of diminishing returns when more than 3 bits are used. Further, we show that using multiple levels for the multiplicative decrease policy enables the protocol to adjust its rate of convergence to fairness, rate variations and responsiveness to congestion based on the degree of congestion at the bottleneck. Based on these fundamental insights, we design multi-level feedback congestion control protocol (MLCP). In addition to being efficient, MLCP converges to a fair bandwidth allocation in the presence of diverse RTT flows while maintaining near-zero packet drop rate and low persistent queue length. These features coupled with MLCP's smooth rate variations make it a viable choice for many real-time applications. Using extensive packet- level simulations we show that the protocol is stable across a diverse range of network scenarios. A fluid model for the protocol shows that MLCP remains globally stable for the case of a single bottleneck link shared by identical round-trip time flows.