Starlink has become the largest commercial LEO satellite network, yet little is known about its internal queue management and bandwidth allocation mechanisms. Prior measurement studies have documented performance variations but lack the granularity to explain the underlying causes. We present the first microscopic characterization of Starlink's transmission behavior, using controlled measurements from multiple terminals to capture per-packet dynamics at microsecond precision. Our analysis uncovers several previously undocumented mechanisms. Starlink employs head-drop queuing rather than tail-drop, with capacities of approximately 1500 and 4000 packets on downlink and uplink, respectively. Bandwidth allocation is demand-driven, starting from a baseline of 100/30 Mbps on the downlink and uplink that ramps up by 3.4×/2× over 400 ms when flows sustain queue pressure. Active queue management aggressively induces packet loss to control queue occupancy, especially on the uplink. These mechanisms reset every 15 seconds during Starlink's reconfiguration cycle. We also find flow-level queuing that isolates latency between concurrent flows while coupling their loss on the downlink. These findings reveal that Starlink's queue management creates fundamentally different operating conditions than terrestrial networks.

Edge-cloud inference collaborations are often designed with a routing estimator that decides whether to offload each frame from weak models at the edge to stronger models in the cloud. Existing systems place the routing estimator after the weak detector, so the weak forward pass still runs even on frames that are later offloaded. In this paper, we argue that this weak-conditioned design can be suboptimal when the offload budget varies. First, we present a competitive weak-skipping estimator (0.153 GFLOPs, about 29x lighter than the weak detector at 4.49 GFLOPs) that extracts routing signal from raw pixels, outperforming the common after-weak placement weak-conditioned baselines. Second, we show that neither weak-skipping nor weak-conditioned placement dominates across the full operating curve, and we propose budget-adaptive routing, which selects between them by offload budget via two offline-tuned thresholds. On PASCAL VOC, our budget-adaptive router traces the upper accuracy envelope of both fixed placements across the operating range. Our method reduces per-frame latency by up to 19.1 ms (about 30% lower at rho = 0.9). Besides outperforming SOTA methods, it is surprisingly stronger than the strong model (+1.7 pp over the strong model's peak mAP) at some operating points with far less compute. Artifacts are available at https://github.com/ViGeng/bgt-ada

Low Earth Orbit (LEO) satellite ISPs promise universal Internet connectivity, yet their interaction with content delivery remains poorly understood. We present the first comprehensive measurement study decomposing Starlink's web content delivery performance decomposed across Point of Presence (PoP), DNS, and CDN layers. Through two years of measurements combining 225K Cloudflare AIM tests, M-Lab data, and active probing from 99 RIPE Atlas and controlled Starlink probes, we collect 6.1M traceroutes and 10.8M DNS queries to quantify how satellite architecture disrupts terrestrial CDN assumptions. We identify three distinct performance regimes based on infrastructure density. Regions with local content-rich PoPs achieve near-terrestrial latencies with the satellite segment dominating 80-90% of RTT. Infrastructure-sparse regions suffer cascading penalties: remote PoPs force distant resolver selection, which triggers CDN mis-localization, pushing latencies beyond 200 ms. Dense-infrastructure regions show minimal sensitivity to PoP changes. Leveraging Starlink's infrastructure expansion in early 2025 as a natural experiment, we demonstrate that relocating PoPs closer to user location reduces median page-fetch times by 60%. Our findings reveal that infrastructure proximity, not satellite coverage, influences web performance, requiring fundamental changes to CDN mapping and DNS resolution for satellite ISPs.

Content Delivery Networks (CDNs) have been pivotal in the dramatic evolution of the Internet, handling the majority of data traffic for billions of connected users. Low-Earth-Orbit (LEO) satellite networks, such as Starlink, aim to revolutionize global connectivity by providing high-speed, low-latency Internet to remote regions. However, LEO satellite networks (LSNs) face challenges integrating with traditional CDNs, which rely on geographical proximity for efficient content delivery - a method that clashes with the operational dynamics of LSNs. In this paper, we scrutinize the operation of CDNs in the context of LSNs, using Starlink as a case study. We develop a browser extension NetMet that performs extensive web browsing experiments from controlled nodes using both Starlink and terrestrial Internet access. Additionally, we analyse crowdsourced speed tests from Starlink users to Cloudflare CDN servers globally. Our results indicate significant performance issues for Starlink users, stemming from the misalignment between terrestrial and satellite infrastructures. We then investigate the potential for SpaceCDNs which integrate CDN infrastructure directly within the LSNs, and show that this approach offers a promising alternative that decreases latencies by over 50%, making them comparable with the CDN experience of users behind terrestrial ISPs. Our aim is to stimulate further research and discussion on overcoming the challenges of effective content delivery with growing LSN offerings.

Low-Earth-Orbit satellite networks (LSNs) are enabling low-latency high-bandwidth internet connectivity at a global scale. However, majority of the traffic on the Internet is currently handled by Content Delivery Networks (CDNs), which rely on geographical proximity to deliver content. In this work, we examine CDN performance for the commercial largest LSN, i.e. Starlink, by performing active measurements through our web browser plugin and passive analysis of Cloudflare speed tests globally. Comparing this to terrestrial networks, we highlight significant performance degradation for Starlink users due to the asymmetries between satellite and terrestrial infrastructure.

The emerging nature of low-earth-orbiting satellite constellations has brought forth an era of global connectivity. Star-link has the potential to offer a truly global, high-performant service. An obvious question arising from this is whether Starlink can replace existing terrestrial services. In this work we describe the key findings from a detailed study of the performance of Starlink, investigating both its potential and shortcomings to act as a "global ISP".

The Starlink network from SpaceX stands out as the only commercial LEO network with over 2M+ customers and more than 4000 operational satellites. In this paper, we conduct a first-of-its-kind extensive multi-faceted analysis of Starlink performance leveraging several measurement sources. First, based on 19.2M crowdsourced M-Lab speed tests from 34 countries since 2021, we analyze Starlink global performance relative to terrestrial cellular networks. Second, we examine Starlink's ability to support real-time latency and bandwidth-critical applications by analyzing the performance of (i) Zoom conferencing, and (ii) Luna cloud gaming, comparing it to 5G and fiber. Third, we perform measurements from Starlink-enabled RIPE Atlas probes to shed light on the last-mile access and other factors affecting its performance.Finally, we conduct controlled experiments from Starlink dishes in two countries and analyze the impact of globally synchronized "15-second reconfiguration intervals'' of the satellite links that cause substantial latency and throughput variations. Our unique analysis paints the most comprehensive picture of Starlink's global and last-mile performance to date.

Widespread adoption of mobile augmented reality (AR) and virtual reality (VR) applications depends on their smoothness and immersiveness. Modern AR applications applying computationally intensive computer vision algorithms can burden today's mobile devices, and cause high energy consumption and/or poor performance. To tackle this challenge, it is possible to offload part of the computation to nearby devices at the edge. However, this calls for smart task placement strategies in order to efficiently use the resources of the edge infrastructure. In this paper, we introduce Nimbus — a task placement and offloading solution for a multi-tier, edge-cloud infrastructure where deep learning tasks are extracted from the AR application pipeline and offloaded to nearby GPU-powered edge devices. Our aim is to minimize the latency experienced by end-users and the energy costs on mobile devices. Our multifaceted evaluation, based on benchmarked performance of AR tasks, shows the efficacy of our solution. Overall, Nimbus reduces the task latency by$\sim 4\times$and the energy consumption by$\sim$77% for real-time object detection in AR applications. We also benchmark three variants of our offloading algorithm, disclosing the trade-off of centralized versus distributed execution.

Emerging Remote Piloting (RP) operations of electrified Unmanned Aerial Vehicles (UAVs) demand low-latency and high-quality video delivery to conduct safe operations in the low-altitude airspace. Although cellular networks are one of the prominent candidates to provide connectivity for such operations, their ground-centric nature limits their capabilities in achieving seamless and reliable aerial connectivity. In this paper, we study the feasibility of supporting RP operations with low latency and high-quality video delivery over commercial cellular networks. By setting up an adaptive bitrate video transmission pipeline with the Google Congestion Control (GCC) and Self-Clocked Rate Adaptation for Multimedia (SCReAM) Congestion Control (CC) algorithms, we analyze the video delivery performance for the RP application requirements and compare the performance of GCC and SCReAM against constant bitrate video delivery. Our results show that low-latency video delivery with < 300 ms playback latency between full-HD and 4K resolution can be maintained up to about 95% of the time in the air. While static bitrate video delivery outperforms adaptive streaming in urban location with abundant link capacity, the latter becomes advantageous in rural locations, where the link capacity is affected by fluctuations. Although the study's findings highlight the capabilities of cellular networks in delivering low-latency video for a safety-critical aerial service, we also discuss the potential improvements and future research challenges for enabling safe operations and meeting the service requirements using cellular networks. We release our collected traces and the video transmission pipeline as open-source to facilitate research in this field.

Cloud computing has seen continuous growth over the last decade. The recent rise in popularity of next-generation applications brings forth the question: "Can current cloud infrastructure support the low latency requirements of such apps?" Specifically, the interplay of wireless last-mile and investments of cloud operators in setting up direct peering agreements with ISPs globally to current cloud reachability and latency has remained largely unexplored.

Multipath TCP (MPTCP) extends traditional TCP to enable simultaneous use of multiple connection endpoints at the source and destination. MPTCP has been under active development since its standardization in 2013, and more recently in February 2020, MPTCP was upstreamed to the Linux kernel. In this paper, we provide the first broad analysis of MPTCPv0 in the Internet. We probe the entire IPv4 address space and an IPv6 hitlist to detect MPTCP-enabled systems operational on port 80 and 443. Our scans reveal a steady increase in MPTCP-capable IPs, reaching 9k+ on IPv4 and a few dozen on IPv6. We also discover a significant share of seemingly MPTCP-capable hosts, an artifact of middleboxes mirroring TCP options. We conduct targeted HTTP(S) measurements towards select hosts and find that middleboxes can aggressively impact the perceived quality of applications utilizing MPTCP. Finally, we analyze two complementary traffic traces from CAIDA and MAWI to shed light on the real-world usage of MPTCP. We find that while MPTCP usage has increased by a factor of 20 over the past few years, its traffic share is still quite low.

Edge computing aims to enable applications with stringent latency requirements, e.g., augmented reality, and tame the overwhelming data streams generated by IoT devices. A core principle of this paradigm is to bring the computation from a distant cloud closer to service consumers and data producers. Consequentially, the issue of edge computing facilities' placement arises. We present a comprehensive analysis suggesting where to place general-purpose edge computing resources on an Internet-wide scale. We base our conclusions on extensive real-world network measurements. We perform extensive traceroute measurements from RIPE Atlas to datacenters in the US, resulting in a graph of 11K routers. We identify the affiliations of the routers to determine the network providers that can act as edge providers. We devise several edge placement strategies and show that they can improve cloud access latency by up to 30%.

In the early days of cloud computing, datacenters were sparsely deployed at distant locations far from end-users with high end-to-end communication latency. However, today’s cloud datacenters have become more geographically spread, the bandwidth of the networks keeps increasing, pushing the end-users latency down. In this paper, we provide a comprehensive cloud reachability study as we perform extensive global client-to-cloud latency measurements towards 189 datacenters from all major cloud providers. We leverage the well-known measurement platform RIPE Atlas, involving up to 8500 probes deployed in heterogeneous environments, e.g., home and offices. Our goal is to evaluate the suitability of modern cloud environments for various current and predicted applications. We achieve this by comparing our latency measurements against known human perception thresholds and are able to draw inferences on the suitability of current clouds for novel applications, such as augmented reality. Our results indicate that the current cloud coverage can easily support several latency-critical applications, like cloud gaming, for the majority of the world’s population.

Edge computing has gained attention from both academia and industry by pursuing two significant challenges: 1) moving latency critical services closer to the users, 2) saving network bandwidth by aggregating large flows before sending them to the cloud. While the rationale appeared sound at its inception almost a decade ago, several current trends are impacting it. Clouds have spread geographically reducing end-user latency, mobile phones? computing capabilities are improving, and network bandwidth at the core keeps increasing. In this paper, we scrutinize edge computing, examining its outlook and future in the context of these trends. We perform extensive client-to-cloud measurements using RIPE Atlas, and show that latency reduction as motivation for edge is not as persuasive as once believed; for most applications the cloud is already 'close enough' for majority of the world's population. This implies that edge computing may only be applicable for certain application niches, as opposed to a general-purpose solution.

AWS offers discounted transient virtual instances as a way to sell unused resources in their data-centers, and users can enjoy up to 90% discount as compared to the regular on-demand pricing. Despite the economic incentives to purchase these transient instances, they do not come with regular availability SLAs, meaning that they can be evicted at any moment. Hence, the user is responsible for managing the instance availability to meet the application requirements. In this paper, we present Bricklayer, a software tool that assists users to better use transient resources in the cloud, reducing costs for the same amount of resources, and increasing the overall instance availability. Bricklayer searches for possible combinations of smaller and cheaper instances to compose the requested amount of resources while deploying them into different spot markets to reduce the risk of eviction. We implemented and evaluated Bricklayer using 3 months of historical data from AWS and found out that it can reduce up 54% of the regular spot price and up to 95% compared to the standard on-demand pricing.

The sharing economy has made great inroads with services like Uber or Airbnb enabling people to share their unused resources with those needing them. The computing world, however, despite its abundance of excess computational resources has remained largely unaffected by this trend, save for few examples like SETI@home. We present DeCloud, a decentralized market framework bringing the sharing economy to on-demand computing where the offering of pay-as-you-go services will not be limited to large companies, but ad hoc clouds can be spontaneously formed on the edge of the network. We design incentive compatible double auction mechanism targeted specifically for distributed ledger trust model instead of relying on third-party auctioneer. DeCloud incorporates innovative matching heuristic capable of coping with the level of heterogeneity inherent for large-scale open systems. Evaluating DeCloud on Google cluster-usage data, we demonstrate that the system has a near-optimal performance from an economic point of view, additionally enhanced by the flexibility of matching.

The Internet is largely a self-organizing system that adapts to changes in its operating environment. In this work, we extend these principles to service infrastructure and introduce ICON, standing for intelligent container. Technically, ICON is a container encapsulating a service that is consumed either directly by end-clients or other services. The novelty of ICON is in the ability of containers to adapt to their environment, targeting near-optimal service delivery and requiring only high-level guidance from the application management. Once deployed, containers form an overlay, observe their setting, and migrate or replicate themselves as needed, to the locations e.g., closest to service consumers. ICON captures our long-term vision for self-organizing service overlays that have the potential for global outreach. Bringing intelligence and adaptation to the level of individual containers renders a decentralized solution that has desirable properties, such as scalability, resilience, reliability, and adaptability to volatile environments. We hope that technology like ICON can open the way for more democratized service provisioning, disintermediating service providers from centralized brokers and optimizing orchestrators.

Edge clouds are an attractive platform to support latency-sensitive applications by providing computations on servers deployed close to end-users. These servers aim to employ MPTCP to leverage multiple connections including wireless over a public network. In this paper, we show that the default MPTCP design does not adequately support reliability in these environments, which makes it unfit for use in edge clouds. We propose RAMPTCP, an extension to MPTCP which focuses on adding reliability over network paths.