As deep learning continues to be driven by ever-larger datasets, understanding which examples are most important for generalization has become a critical question. While progress in data selection continues, emerging applications require studying this problem in dynamic contexts. To bridge this gap, we pose the Incremental Data Selection (IDS) problem, where examples arrive as a continuous stream, and need to be selected without access to the full data source. In this setting, the learner must incrementally build a training dataset of predefined size while simultaneously learning the underlying task. We find that in IDS, the impact of a new sample on the model state depends fundamentally on both its geometric relationship in the feature space and its prediction error. Leveraging this insight, we propose PEAKS (Prediction Error Anchored by Kernel Similarity), an efficient data selection method tailored for IDS. Our comprehensive evaluations demonstrate that PEAKS consistently outperforms existing selection strategies. Furthermore, PEAKS yields increasingly better performance returns than random selection as training data size grows on real-world datasets. The code is available at https://github.com/BurakGurbuz97/PEAKS.

Deep neural networks (DNNs) struggle to learn in dy-namic settings because they mainly rely on static datasets. Continual learning (CL) aims to overcome this limitation by enabling DNNs to incrementally accumulate knowledge. A widely adopted scenario in CL is class-incremental learning (CIL), where DNNs are required to sequentially learn more classes. Among the various strategies in CL, replay methods, which revisit previous classes, stand out as the only effective ones in CIL. Other strategies, such as ar-chitectural modifications to segregate information across weights and protect them from change, are ineffective in CIL. This is because they need additional information during testing to select the correct network parts to use. In this paper, we propose NICE, Neurogenesis Inspired Contex-tual Encoding, a replay-free architectural method inspired by adult neurogenesis in the hippocampus. NICE groups neurons in the DNN based on different maturation stages and infers which neurons to use during testing without any additional signal. Through extensive experiments across 6 datasets and 3 architectures, we show that NICE performs on par with or often outperforms replay methods. We also make the case that neurons exhibit highly distinctive acti-vation patterns for the classes in which they specialize, en-abling us to determine when they should be used. The code is available at https://github.com/BurakGurbuz97/NICE.

Natural target functions and tasks typically exhibit hierarchical modularity -- they can be broken down into simpler sub-functions that are organized in a hierarchy. Such sub-functions have two important features: they have a distinct set of inputs (input-separability) and they are reused as inputs higher in the hierarchy (reusability). Previous studies have established that hierarchically modular neural networks, which are inherently sparse, offer benefits such as learning efficiency, generalization, multi-task learning, and transfer. However, identifying the underlying sub-functions and their hierarchical structure for a given task can be challenging. The high-level question in this work is: if we learn a task using a sufficiently deep neural network, how can we uncover the underlying hierarchy of sub-functions in that task? As a starting point, we examine the domain of Boolean functions, where it is easier to determine whether a task is hierarchically modular. We propose an approach based on iterative unit and edge pruning (during training), combined with network analysis for module detection and hierarchy inference. Finally, we demonstrate that this method can uncover the hierarchical modularity of a wide range of Boolean functions and two vision tasks based on the MNIST digits dataset.

The goal of continual learning (CL) is to learn different tasks over time. The main desiderata associated with CL are to maintain performance on older tasks, leverage the latter to improve learning of future tasks, and to introduce minimal overhead in the training process (for instance, to not require a growing model or retraining). We propose the Neuro-Inspired Stability-Plasticity Adaptation (NISPA) architecture that addresses these desiderata through a sparse neural network with fixed density. NISPA forms stable paths to preserve learned knowledge from older tasks. Also, NISPA uses connection rewiring to create new plastic paths that reuse existing knowledge on novel tasks. Our extensive evaluation on EMNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets shows that NISPA significantly outperforms representative state-of-the-art continual learning baselines, and it uses up to ten times fewer learnable parameters compared to baselines. We also make the case that sparsity is an essential ingredient for continual learning. The NISPA code is available at https://github.com/BurakGurbuz97/NISPA.

Methods that sparsify a network at initialization are important in practice because they greatly improve the efficiency of both learning and inference. Our work is based on a recently proposed decomposition of the Neural Tangent Kernel (NTK) that has decoupled the dynamics of the training process into a data-dependent component and an architecture-dependent kernel {–} the latter referred to as Path Kernel. That work has shown how to design sparse neural networks for faster convergence, without any training data, using the Synflow-L2 algorithm. We first show that even though Synflow-L2 is optimal in terms of convergence, for a given network density, it results in sub-networks with “bottleneck” (narrow) layers {–} leading to poor performance as compared to other data-agnostic methods that use the same number of parameters. Then we propose a new method to construct sparse networks, without any training data, referred to as Paths with Higher-Edge Weights (PHEW). PHEW is a probabilistic network formation method based on biased random walks that only depends on the initial weights. It has similar path kernel properties as Synflow-L2 but it generates much wider layers, resulting in better generalization and performance. PHEW achieves significant improvements over the data-independent SynFlow and SynFlow-L2 methods at a wide range of network densities.

We first pose the Unsupervised Progressive Learning (UPL) problem: an online representation learning problem in which the learner observes a non-stationary and unlabeled data stream, learning a growing number of features that persist over time even though the data is not stored or replayed. To solve the UPL problem we propose the Self-Taught Associative Memory (STAM) architecture. Layered hierarchies of STAM modules learn based on a combination of online clustering, novelty detection, forgetting outliers, and storing only prototypical features rather than specific examples. We evaluate STAM representations using clustering and classification tasks. While there are no existing learning scenarios that are directly comparable to UPL, we compare the STAM architecture with two recent continual learning models, Memory Aware Synapses (MAS) and Gradient Episodic Memories (GEM), after adapting them in the UPL setting.

We approach the C. elegans connectome as an information processing network that receives input from about 90 sensory neurons, processes that information through a highly recurrent network of about 80 interneurons, and it produces a coordinated output from about 120 motor neurons that control the nematode's muscles. We focus on the feedforward flow of information from sensory neurons to motor neurons, and apply a recently developed network analysis framework referred to as the "hourglass effect". The analysis reveals that this feedforward flow traverses a small core ("hourglass waist") that consists of 10-15 interneurons. These are mostly the same interneurons that were previously shown (using a different analytical approach) to constitute the "rich-club" of the C. elegans connectome. This result is robust to the methodology that separates the feedforward from the feedback flow of information. The set of core interneurons remains mostly the same when we consider only chemical synapses or the combination of chemical synapses and gap junctions. The hourglass organization of the connectome suggests that C. elegans has some similarities with encoder-decoder artificial neural networks in which the input is first compressed and integrated in a low-dimensional latent space that encodes the given data in a more efficient manner, followed by a decoding network through which intermediate-level sub-functions are combined in different ways to compute the correlated outputs of the network. The core neurons at the hourglass waist represent the information bottleneck of the system, balancing the representation accuracy and compactness (complexity) of the given sensory information.

The development and adoption of new protocols (or of extensions to existing protocols) is arguably central to the Internet's evolution. However, and in spite of over 40 years of experience with this process, we have limited understanding of what factors may contribute to a protocol's success. A sound technical design and a well-grounded purpose are obviously important, but we have many examples of failures that met those two criteria. What other factors affect a protocol's likelihood of success, and under what circumstances? We investigate this question through a statistical approach, based on analyzing a set of about 250 Internet standard documents, Internet engineering task force request for comments (RFCs). We characterize these RFCs using a number of key features, which we then seek to associate with positive or negative odds when it comes to success. Our high-level results are intuitive, e.g., protocols that call for Internet-wide adoption face greater challenges. Focusing on more targeted subsets of protocols reveals more subtle and possibly more interesting differences between areas of the Internet landscape. We also apply our prediction framework to IPv6, and use different “what-if” scenarios to explore what might have affected its deployment.

Data represented as strings abounds in biology, linguistics, document mining, web search and many other fields. Such data often have a hierarchical structure, either because they were artificially designed and composed in a hierarchical manner or because there is an underlying evolutionary process that creates repeatedly more complex strings from simpler substrings. We propose a framework, referred to as Lexis, that produces an optimized hierarchical representation of a given set of "target" strings. The resulting hierarchy, "Lexis-DAG", shows how to construct each target through the concatenation of intermediate substrings, minimizing the total number of such concatenations or DAG edges. The Lexis optimization problem is related to the smallest grammar problem. After we prove its NP-hardness for two cost formulations, we propose an efficient greedy algorithm for the construction of Lexis-DAGs. We also consider the problem of identifying the set of intermediate nodes (substrings) that collectively form the "core" of a Lexis-DAG, which is important in the analysis of Lexis-DAGs. We show that the Lexis framework can be applied in diverse applications such as optimized synthesis of DNA fragments in genomic libraries, hierarchical structure discovery in protein sequences, dictionary-based text compression, and feature extraction from a set of documents.

What determines the eventual success of a protocol? Are certain features or properties more important? Do those vary according to a protocol's type? We explore these questions by applying data mining techniques to a rich repository of protocol specifications; IETF RFCs. While the investigation is still preliminary, some interesting findings have emerged. It confirms a number of intuitive results such as backward compatibility being key for protocol extensions and new versions, but not for new protocols. Similarly, the ability to improve performance is the single most important factor in the success of data plane protocols. Less intuitive findings, however, also emerge. Adding value to other protocols was the most significant factor in the success of new protocols, while extensions targeting security were the most likely to fail among new application and transport layer protocols. The paper offers a brief overview of our methodology and of the initial results it has afforded.

Peering in the Internet interdomain network has long been considered a “black art”, understood in-depth only by a select few peering experts while the majority of the network operator community only scratches the surface employing conventional rules-of-thumb to form peering links through ad hoc personal interactions. Why is peering considered a black art? What are the main sources of complexity in identifying potential peers, negotiating a stable peering relationship, and utility optimization through peering? How do contemporary operational practices approach these problems? In this work we address these questions for Tier-2 Network Service Providers. We identify and explore three major sources of complexity in peering: (a) inability to predict traffic flows prior to link formation (b) inability to predict economic utility owing to a complex transit and peering pricing structure (c) computational infeasibility of identifying the optimal set of peers because of the network structure. We show that framing optimal peer selection as a formal optimization problem and solving it is rendered infeasible by the nature of these problems. Our results for traffic complexity show that 15% NSPs lose some fraction of customer traffic after peering. Additionally, our results for economic complexity show that 15% NSPs lose utility after peering, approximately, 50% NSPs end up with higher cumulative costs with peering than transit only, and only 10% NSPs get paid-peering customers.

Content delivery networks (CDNs) deploy globally distributed systems of caches in a large number of autonomous systems (ASes). It is important for a CDN operator to satisfy the performance requirements of end users, while minimizing the cache deployment cost. In this paper, we study the cache deployment optimization (CaDeOp) problem of determining how much server, energy, and bandwidth resources to provision in each cache AS, i.e., each AS chosen for cache deployment. The CaDeOp objective is to minimize the total cost incurred by the CDN, subject to meeting the end-user performance requirements. We formulate the CaDeOp problem as a mixed integer program (MIP) and solve it for realistic AS-level topologies, traffic demands, and non-linear energy and bandwidth costs. We also evaluate the sensitivity of the results to our parametric assumptions. When the end-user performance requirements become more stringent, the CDN footprint rapidly expands, requiring cache deployments in additional ASes and geographical regions. Also, the CDN cost increases several times, with the cost balance shifting toward bandwidth and energy costs. On the other hand, the traffic distribution among the cache ASes stays relatively even, with the top 20% of the cache ASes serving around 30% of the overall traffic.

Peering agreements between Autonomous Systems affect not only the flow of interdomain traffic but also the economics of the entire Internet ecosystem. The conventional wisdom is that transit providers are selective in choosing their settlement-free peers because they prefer to offer revenue-generating transit service to others. Surprisingly, however, a large percentage of transit providers use an Open peering strategy. What causes this large-scale adoption of Open peering, especially among transit providers? More importantly, what is the impact of this peering trend on the economic performance of the population of transit providers? We approach these questions through game-theoretic modeling and agent-based simulations, capturing the dynamics of peering strategy adoption, inter-network formation and interdomain traffic flow. We explain why transit providers gravitate towards Open peering even though that move may be detrimental to their economic fitness. Finally, we examine the impact of an Open peering variant that requires some coordination among providers.

Video streaming is a major source of Internet traffic today and usage continues to grow at a rapid rate. To cope with this new and massive source of traffic, ISPs use methods such as caching to reduce the amount of traffic traversing their networks and serve customers better. However, the presence of a standard cache server in the video transfer path may result in bitrate oscillations and sudden rate changes for Dynamic Adaptive Streaming over HTTP (DASH) clients. In this paper, we investigate the interactions between a client and a cache that result in these problems, and propose an approach to solve it. By adaptively controlling the rate at which the client downloads video segments from the cache, we can ensure that clients will get smooth video. We verify our results using simulation and show that compared to a standard cache our approach (1) can reduce bitrate oscillations (2) prevents sudden rate changes, and compared to a no-cache scenario (3) provides traffic savings, and (4) improves the quality of experience of clients.

Prior work has shown that two or more adaptive streaming players can be unstable when they compete for bandwidth. The root cause of the instability problem is that, in Steady-State, a player goes through an ON-OFF activity pattern in which it overestimates the available bandwidth. We propose a server-based traffic shaping method that can significantly reduce such oscillations without significant (or any) loss in bandwidth utilization. The shaper is only activated when oscillations are detected, and it dynamically adjusts the shaping rate so that the player should ideally receive the highest available video profile while being stable. We evaluate the proposed method experimentally in terms of instability and utilization comparing with the unshaped case, under several scenarios.

To cope with the phenomenal growth of the Internet over the next decade, core networks are expected to scale to capacities of terabits-per-second and beyond. Increasing the role of optics for switching and transmission inside the core network seems to be the most promising way forward to accomplish this capacity scaling. Unfortunately, unlike electronic memory, it remains a formidable challenge to build even a few packets of integrated all-optical buffers. In this context, we envision a bufferless (or near-zero buffer) core optical network and make three contributions: First, we propose a novel edge-to-edge based packet-level forward error correction (FEC) scheme that combats packet loss in the bufferless core, and characterise the impact of FEC strength on loss at a single link. Second, we develop a global optimisation framework for multi-hop networks, and propose a heuristic algorithm that adjusts FEC strength to achieve fairness amongst the different single- and multi-hop flows. Finally, we evaluate the performance of our FEC scheme for realistic mixes of short- and long-lived TCP flows, and show that edge-to-edge packet-level FEC can be tuned to effectively mitigate contention losses in the core, thus opening the doors to bufferless optical networks in the near future.

The focus of network monitoring and problem diagnosis gradually moves towards "soft failures" and performance problems, such as noticeable jitter or loss rate. A major, and still unresolved, problem in this space is to localize a performance problem at the granularity of individual IP-layer links. We propose a network tomography framework that aims to localize performance problems in the set of paths that interconnect a full-mesh topology formed by N sensors. As opposed to Boolean tomography, which models links as "good" versus "bad", we consider a multi-level discrete tomography framework that can capture the large diversity of performance levels seen at the Internet. We evaluate two tomography methods: the first assumes that the performance of a path is determined by the lowest-performance link in the path; the second also takes into account the number of such links. We propose an efficient algorithm for each method and then evaluate their accuracy and also show their superiority over two traditional approaches (Boolean and Analogue) on real single-provider networks (ESnet and Internet2) as well as in an interdomain context (PlanetLab).

The objective of early network tomography approaches was to produce a point estimate for the performance of each network link (Analog tomography). When it became clear that the previous approach is error-prone in practice, research shifted to Boolean tomography where each link is estimated as either "good" or "bad". The Boolean approach is more practical but its resolution is too coarse. We propose a new tomography framework that combines the best of both worlds: we still distinguish between good and bad links (for practicality reasons) but we also infer a range estimate for the performance of each bad link. We apply the Range tomography framework in two path performance metric functions (Min and Sum) and propose an efficient algorithm for each problem. Together with simulations, we have also applied Range tomography in three operational networks allowing us to identify the location of bad links and to estimate their performance during congestion episodes. We also compare the proposed method with existing Analog and Boolean tomography algorithms.

We propose an agent-based network formation model for the Internet at the Autonomous System (AS) level. The proposed model, called GENESIS, is based on realistic provider and peering strategies, with ASes acting in a myopic and decentralized manner to optimize a cost-related fitness function. GENESIS captures key factors that affect the network formation dynamics: highly skewed traffic matrix, policy-based routing, geographic co-location constraints, and the costs of transit/peering agreements. As opposed to analytical game-theoretic models, which focus on proving the existence of equilibria, GENESIS is a computational model that simulates the network formation process and allows us to actually compute distinct equilibria (i.e., networks) and to also examine the behavior of sample paths that do not converge. We find that such oscillatory sample paths occur in about 10% of the runs, and they always involve tier- 1 ASes, resembling the tier-1 peering disputes often seen in practice. GENESIS results in many distinct equilibria that are highly sensitive to initial conditions and the order in which ASes (agents) act. This implies that we cannot predict the properties of an individual AS in the Internet. However, certain properties of the global network or of certain classes of ASes are predictable. We also examine whether the underlying game is zero-sum, and identify three sufficient conditions for that property. Finally, we apply GENESIS in a specific “what-if” question, asking how the openness towards peering affects the resulting network in terms of topology, traffic flow and economics. Interestingly, we find that the peering openness that maximizes the fitness of different network classes (tier-1, tier-2 and tier-3 providers) closely matches that seen in real-world peering policies.

With an increasing demand for high-quality video content over the Internet, it is becoming more likely that two or more adaptive streaming players share the same network bottleneck and compete for available bandwidth. This competition can lead to three performance problems: player instability, unfairness between players, and bandwidth underutilization. However, the dynamics of such competition and the root cause for the previous three problems are not yet well understood. In this paper, we focus on the problem of competing video players and describe how the typical behavior of an adaptive streaming player in its Steady-State, which includes periods of activity followed by periods of inactivity (ON-OFF periods), is the main root cause behind the problems listed above. We use two adaptive players to experimentally showcase these issues. Then, focusing on the issue of player instability, we test how several factors (the ON-OFF durations, the available bandwidth and its relation to available bitrates, and the number of competing players) affect stability.

No abstract available.

The scalability limitations of BGP have been a major concern lately. An important aspect of this issue is the rate of routing updates (churn) that BGP routers must process. This paper presents an analysis of the evolution of churn in four networks at the backbone of the Internet over a period of seven years and eight months, using BGP update traces from the RouteViews project. The churn rate varies widely over time and between networks. Instead of descriptive “black-box” statistical analysis, we take an exploratory data analysis approach attempting to understand the reasons behind major observed characteristics of the churn time series. We find that duplicate announcements are a major churn contributor, responsible for most large spikes. Remaining spikes are mostly caused by routing incidents that affect a large number of prefixes simultaneously. More long-term intense periods of churn, on the other hand, are caused by misconfigurations or other special events at or close to the monitored autonomous system (AS). After filtering pathologies and effects that are not related to the long-term evolution of churn, we analyze the remaining “baseline” churn and find that it is increasing at a rate that is similar to the growth of the number of ASs.

We present an end-to-end measurement method for the detection of traffic shaping. Traffic shaping is typically implemented using token buckets, allowing a maximum burst of traffic to be serviced at the peak capacity of the link, while any remaining traffic is serviced at a lower shaping rate. The contribution of this paper is twofold. First, we develop an active end-to-end detection mechanism, referred to as ShaperProbe, that can infer whether a particular path is subject to traffic shaping, and in that case, estimate the shaper characteristics. Second, we analyze results from a large-scale deployment of ShaperProbe on M-Lab over the last 24 months, detecting traffic shaping in several major ISPs. Our deployment has received more than one million runs so far from 5,700 ISPs.

Adaptive (video) streaming over HTTP is gradually being adopted, as it offers significant advantages in terms of both user-perceived quality and resource utilization for content and network service providers. In this paper, we focus on the rate-adaptation mechanisms of adaptive streaming and experimentally evaluate two major commercial players (Smooth Streaming, Netflix) and one open source player (OSMF). Our experiments cover three important operating conditions. First, how does an adaptive video player react to either persistent or short-term changes in the underlying network available bandwidth. Can the player quickly converge to the maximum sustainable bitrate? Second, what happens when two adaptive video players compete for available bandwidth in the bottleneck link? Can they share the resources in a stable and fair manner? And third, how does adaptive streaming perform with live content? Is the player able to sustain a short playback delay? We identify major differences between the three players, and significant inefficiencies in each of them.

The Internet protocol stack has a layered architecture that resembles an hourglass. The lower and higher layers tend to see frequent innovations, while the protocols at the waist of the hourglass appear to be "ossified". We propose EvoArch, an abstract model for studying protocol stacks and their evolution. EvoArch is based on a few principles about layered network architectures and their evolution in a competitive environment where protocols acquire value based on their higher layer applications and compete with other protocols at the same layer. EvoArch produces an hourglass structure that is similar to the Internet architecture from general initial conditions and in a robust manner. It also suggests a plausible explanation why some protocols, such as TCP or IP, managed to survive much longer than most other protocols at the same layers. Furthermore, it suggests ways to design more competitive new protocols and more evolvable future Internet architectures.

Our goal is to understand the evolution of the autonomous system (AS) ecosystem over the last 12 years. Instead of focusing on abstract topological properties, we classify ASs into a number of types depending on their function and business type. Furthermore, we consider the semantics of inter-AS links: customer-provider versus peering relations. We find that the available historic datasets from RouteViews and RIPE are not sufficient to infer the evolution of peering links, and so we restrict our focus to customer-provider links. Our findings highlight some important trends in the evolution of the Internet over the last 12 years and hint at what the Internet is heading toward. After an exponential increase phase until 2001, the Internet has settled into a slower exponential growth in terms of both ASs and inter-AS links. The growth is mostly due to enterprise networks and content/access providers at the periphery of the Internet. The average path length remains almost constant, mostly due to the increasing multihoming degree of transit and content/access providers. The AS types differ significantly from each other with respect to their rewiring activity; content/access providers are the most active. A few large transit providers act as “attractors” or “repellers” of customers. For many providers, strong attractiveness precedes strong repulsiveness by 3-9 months. Finally, in terms of regional growth, we find that the AS ecosystem is now larger and more dynamic in Europe than in North America.

Recent measurements and anecdotal evidence indicate that the Internet ecosystem is rapidly evolving from a multi-tier hierarchy built mostly with transit (customer-provider) links to a dense mesh formed with mostly peering links. This transition can have major impact on the global Internet economy as well as on the traffic flow and topological structure of the Internet. In this paper, we study this evolutionary transition with an agent-based network formation model that captures key aspects of the interdomain ecosystem, viz., interdomain traffic flow and routing, provider and peer selection strategies, geographical constraints, and the economics of transit and peering interconnections. The model predicts several substantial differences between the Hierarchical Internet and the Flat Internet in terms of topological structure, path lengths, interdomain traffic flow, and the profitability of transit providers. We also quantify the effect of the three factors driving this evolutionary transition. Finally, we examine a hypothetical scenario in which a large content provider produces more than half of the total Internet traffic.

The scalability limitations of BGP have been a major concern in the networking community lately. An important issue in this respect is the rate of routing updates (churn) that BGP routers must process. This paper presents an analysis of the evolution of churn in four networks in the backbone of the Internet over the last six years, using update traces from the Routeviews project. The churn rate varies widely over time and between networks, and cannot be understood through "black-box'' statistical analysis. Instead we take a different approach with a focus on investigating the underlying reasons for BGP churn evolution. Through our analysis we are able to identify and isolate the main reasons behind many of the anomalies in the churn time series. We find that duplicate announcements is a major churn contributor, and responsible for most large spikes in the churn time series. Other intense periods of churn are caused by misconfigurations or other special events in or close to the monitored AS, and hence limiting these is an important mean to limit churn. We then analyze the remaining "baseline'' churn, and find that it is increasing with a rate much slower than the increase in the routing table size.

We propose an active probing method, called Differential Probing or DiffProbe, to detect whether an access ISP is deploying forwarding mechanisms such as priority scheduling, variations of WFQ, or WRED to discriminate against some of its customer flows. DiffProbe aims to detect if the ISP is doing one or both of delay discrimination and loss discrimination. The basic idea in DiffProbe is to compare the delays and packet losses experienced by two flows: an Application flow A and a Probing flow P. The paper describes the statistical methods that DiffProbe uses, a novel method for distinguishing between Strict Priority and WFQ-variant packet scheduling, simulation and emulation experiments, and a few real-world tests at major access ISPs.

Internet traffic is expected to grow phenomenally over the next five to ten years, and to cope with such large traffic volumes, core networks are expected to scale to capacities of terabits-per-second and beyond. Increasing the role of optics for switching and transmission inside the core network seems to be the most promising way forward to accomplish this capacity scaling. Unfortunately, unlike electronic memory, it remains a formidable challenge to build even a few packets of integrated all-optical buffers. In the context of envisioning a bufferless (or near-zero buffer) core network, our contributions are threefold: First, we propose a novel edge-to-edge based packet-level forward error correction (FEC) framework as a means of combating high core losses, and investigate via analysis and simulation the appropriate FEC strength for a single core link. Second, we consider a realistic multi-hop network and develop an optimisation framework that adjusts the FEC strength on a per-flow basis to ensure fairness between single-and multi-hop flows. Third, we study the efficacy of FEC for various system parameters such as relative mixes of short-lived and long-lived TCP flows, and average offered link loads. Our study is the first to show that packet-level FEC, when tuned properly, can be very effective in mitigating high core losses, thus opening the doors to a bufferless core network in the future.

The scalability of BGP routing is a major concern for the Internet community. Scalability is an issue in two different aspects: increasing routing table size, and increasing rate of BGP updates. In this paper, we focus on the latter. Our objective is to characterize the churn increase experienced by ASes in different levels of the Internet hierarchy as the network grows. We look at several "what-if" growth scenarios that are either plausible directions in the evolution of the Internet or educational corner cases, and investigate their scalability implications and interaction with different failure types. Our findings explain the dramatically different impact of multihoming and peering on BGP scalability, highlight negative and positive effects of multihoming on churn and reachability, and identify which topological growth scenarios will lead to faster churn increase for different failure types.

We propose a routing and load-balancing approach with the primary goal of being robust to sudden topological changes and significant traffic matrix variations. The proposed method load-balances traffic over several routes in an adaptive way based on its local view of the load in the network. The focus is on robustness and simplicity, rather than optimality, and so it does not rely on a given traffic matrix, nor it is tuned to a specific topology. Instead, we aim to achieve a satisfactory routing under a wide range of traffic and topology scenarios based on each node's independent operation. The scheme avoids the instability risks of previous load-responsive routing schemes, it does not load the control plane with congestion-related signaling, and it can be implemented on top of existing routing protocols. In this paper, we present the proposed scheme, discuss how it aims to meet the objectives of robustness and load-responsiveness, and evaluate its performance under diverse traffic loads and topological changes with flow-level simulations.

File sharing using peer-to-peer (p2p) systems is a major Internet application and the leading source of network traffic today. However, the dominance of p2p systems for file sharing has been recently challenged by an increasing number of services, such as RapidShare and MegaUpload, which offer users the ability to share files through centralized servers, without relying on an underlying p2p infrastructure. These services, referred to as One-Click Hosting (OCH), have the potential to offer users better performance and availability than p2p systems. If they succeed, OCH services may become the leading platform for file sharing and eventually replace p2p systems for this purpose. In this paper, we present the first, to our knowledge, detailed study of OCH traffic and services focusing on the most popular such service: RapidShare. Through a combination of passive and active measurements, we attempt to understand their service architecture, usage patterns, and content characteristics. We also compare RapidShare with BitTorrent in terms of user-perceived throughput and content availability, and we explore the characteristics of some popular RapidShare indexing sites.

The scalability of BGP routing is a major concern for the Internet community. Scalability is an issue in two different aspects: increasing routing table size, and increasing rate of BGP updates. In this paper, we focus on the latter. Our objective is to characterize the churn increase experienced by ASes in different levels of the Internet hierarchy as the network grows. We look at several what-if growth scenarios that are either plausible directions in the evolution of the Internet or educational corner cases, and investigate their scalability implications. In addition, we examine the effect of the BGP update rate-limiting timer (MRAI), considering both major variations with which it has been deployed. Our findings explain the dramatically different impact of multi-homing and peering on BGP scalability, identify which topological growth scenarios will lead to faster churn increase, and emphasize the importance of not rate-limiting explicit withdrawals (despite what RFC-4271 recently required).

Our goal is to understand the evolution of the Autonomous System (AS) ecosystem over the last decade. Instead of focusing on abstract topological properties, we classify ASes into a number of "species" depending on their function and business type. Further, we consider the semantics of inter-AS links, in terms of customer-provider versus peering relations. We find that the available historic datasets from RouteViews and RIPE are not sufficient to infer the evolution of peering links, and so we restrict our focus to customer-provider links. Our findings highlight some important trends in the evolution of the Internet over the last decade, and hint at what the Internet is heading towards. After an exponential increase phase until 2001, the Internet now grows linearly in terms of both ASes and inter-AS links. The growth is mostly due to enterprise networks and content/access providers at the periphery of the Internet. The average path length remains almost constant mostly due to the increasing multihoming degree of transit and content/access providers. In recent years, enterprise networks prefer to connect to small transit providers, while content/access providers connect equally to both large and small transit providers. The AS species differ significantly from each other with respect to their rewiring activity; content/access providers are the most active. A few large transit providers act as "attractors" or "repellers" of customers. For many providers, strong attractiveness precedes strong repulsiveness by 3-9 months. Finally, in terms of regional growth, we find that the AS ecosystem is now larger and more dynamic in Europe than in North America.

We present an end-to-end active probing methodology that creates frequency-domain signals in IP network paths. The signals are generated by periodic packet trains that cause short-lived queueing delay spikes. Different probers can be multiplexed in the frequency-domain on the same path. Further, a signal that is introduced by a "prober" in one path can cause a crosstalk effect, inducing a signal of the same frequency into another path (the "sampler") as long as the two paths share one or more bottleneck queues. Applications of the proposed methodology include the detection of shared store-and-forward devices among two or more paths, the creation of covert channels, and the modulation of voice or video periodic packet streams in less noisy frequencies. In this paper we focus on the first application. Our goal is to detect shared bottleneck(s) between a "sampler" and one or more "prober" paths. We present a spectral probing methodology as well as the corresponding signal processing/detection process. The accuracy of the method has been evaluated with controlled and repeatable simulation experiments, and it has also been tested on some Internet paths.

The distributed nature of the Internet makes it difficult for a single service provider to troubleshoot the disruptions experienced by its customers. We propose NetDiagnoser, a troubleshooting algorithm to identify the location of failures in an internetwork environment. First, we adapt the well-known Boolean tomography technique to work in this environment. Then, we significantly extend this technique to improve the diagnosis accuracy in the presence of multiple link failures, logical failures (for instance, misconfigurations of route export filters), and incomplete topology inference. In particular, NetDiagnoser takes advantage of rerouted paths, routing messages collected at one provider's network and Looking Glass servers. We evaluate each feature of Net-Diagnoser separately using C-BGP simulations on realistic topologies. Our results show that NetDiagnoser can successfully identify a small set of links, which almost always includes the actually failed/misconfigured links.

The issue of router buffer sizing is still open and significant. Previous work either considers open-loop traffic or only analyzes persistent TCP flows. This paper differs in two ways. First, it considers the more realistic case of non-persistent TCP flows with heavy-tailed size distribution. Second, instead of only looking at link metrics, we focus on the impact of buffer sizing on TCP performance. Specifically, our goal is to find the buffer size that maximizes the average per-flow TCP throughput. Through a combination of testbed experiments, simulation, and analysis, we reach the following conclusions. The output/input capacity ratio at a network link largely determines the required buffer size. If that ratio is larger than one, the loss rate drops exponentially with the buffer size and the optimal buffer size is close to zero. Otherwise, if the output/input capacity ratio is lower than one, the loss rate follows a power-law reduction with the buffer size and significant buffering is needed, especially with flows that are mostly in congestion-avoidance. Smaller transfers, which are mostly in slow-start, require significantly smaller buffers. We conclude by revisiting the ongoing debate on "small versus large" buffers from a new perspective.

Multihoming and overlay routing are used, mostly separately, to bypass Internet outages, congested links and long routes. In this paper, we examine a scenario in which multihoming and overlay routing are jointly used. Specifically, we assume that an overlay service provider (OSP) aims to offer its customers the combined benefits of multihoming and overlay routing, in terms of improved performance, availability and reduced cost, through a network of multihomed overlay routers. We focus on the corresponding design problem, i.e., where to place the overlay routers and how to select the upstream ISPs for each router, with the objective to maximize the profit of the OSP. We examine, with realistic network performance and pricing data, whether the OSP can provide a network service that is profitable, better (in terms of round-trip time), and less expensive than the competing native ISPs. Perhaps surprisingly, we find out that the OSP can meet all three objectives at the same time. We also show that the MON design process is crucial. For example, operating more than 10 overlay nodes or routing traffic through the minimum-delay overlay path, rarely leads to profitability in our simulations.

Multihoming has been used by stub networks for several years as a form of redundancy, improving the availability of Internet access. More recently, Intelligent Route Control (IRC) products allow multihomed networks to dynamically switch parts of their egress or ingress traffic between ISPs, also improving cost and performance. IRC products assume that the set of upstream ISPs is given and fixed. Typically, however, a multihomed network has several ISP choices and the actual selection of ISPs can significantly affect cost, availability, and performance. In the first part of this work, we develop a methodology to select the best set of upstream ISPs, optimizing monetary cost and availability. Our results, based on measurements of actual Internet traffic and topology, show that the proposed algorithm selects the best possible set of ISPs in terms of resiliency to inter-AS single-link failures. The algorithm also performs well in the presence of double or triple link failures. In the second part of this work, we focus on the egress path selection problem. Specifically, we propose a stochastic search algorithm, based on simulated annealing, to allocate the network’s egress traffic between upstream ISPs. The objectives are to minimize cost, also ensuring that the selected paths to the major destinations of egress traffic are congestion-free. Simulation results show that the proposed algorithm performs very well in meeting the previous objectives, when congestion-free paths exist.

Abstract — Intelligent Route Control (IRC) systems are increasingly deployed in multihomed networks. IRC systems aim to optimize the cost and performance of outgoing traffic, based on measurement-driven dynamic path switching techniques. In this paper, we first show that IRC systems can introduce sustained traffic oscillations, causing significant performance degradation instead of improvement. This happens, first, when IRC systems do not take into account the self-load effect, i.e., when they ignore that the performance of a path can change after additional traffic is switched to that path. Second, oscillations can take place when different IRC systems get synchronized due to significant overlap of their measurement time windows. We then propose measurement methodologies and path switching algorithms that can effectively deal with the previous two issues. The proposed IRC techniques use available bandwidth estimation to avoid the self-load effect, and they introduce a random component in the path switching decision or time scale. We evaluate the proposed techniques under diverse traffic conditions. When the background traffic is stationary, IRC systems should switch paths conservatively, only upon major traffic fluctuations. With nonstationary background traffic and congestion periods that last for a time scale ¢¡, IRC systems improve performance only if they can detect congestion and switch paths much faster than ¡; otherwise, they cause oscillations and hurt performance. We also show that the gradual deployment of randomized IRC systems, in the presence of traffic from deterministic IRC systems, can play a stabilizing role and benefits early adopters.

Packet buffers in router/switch interfaces constitute a central element of packet networks. The appropriate sizing of these buffers is an important and open research problem. Much of the previous work on buffer sizing modeled the traffic as an exogenous process, i.e., independent of the network state, ignoring the fact that the offered load from TCP flows depends on delays and losses in the network. In TCP-aware work, the objective has often been to maximize the utilization of the link, without considering the resulting loss rate. Also, previous TCP-aware buffer sizing schemes did not distinguish between flows that are bottlenecked at the given link and flows that are bottlenecked elsewhere, or that are limited by their size or advertised window. In this work, we derive the minimum buffer requirement for a drop-tail link, given constraints on the minimum utilization, maximum loss rate, and maximum queueing delay, when it is feasible to achieve all three constraints. Our results are applicable when most of the traffic (80-90%) at the given link is generated by large TCP flows that are bottlenecked at that link. For heterogeneous flows, we show that the buffer requirement depends on the harmonic mean of their round-trip times, and on the degree of loss synchronization. To limit the maximum loss rate, the buffer should be proportional to the number of flows that are bottlenecked at that link, when that number exceeds a certain threshold. The maximum queueing delay constraint, on the other hand, provides a simple upper bound on the buffer requirement. We also describe how to estimate the parameters of our buffer sizing formula from packet and loss traces, evaluate the proposed model with simulations, and compare it with two other buffer provisioning schemes.

Predicting the throughput of large TCP transfers is important for a broad class of applications. This paper focuses on the design, empirical evaluation, and analysis of TCP throughput predictors. We first classify TCP throughput prediction techniques into two categories: Formula-Based (FB) and History-Based (HB). Within each class, we develop representative prediction algorithms, which we then evaluate empirically over the RON testbed. FB prediction relies on mathematical models that express the TCP throughput as a function of the characteristics of the underlying network path. It does not rely on previous TCP transfers in the given path, and it can be performed with non-intrusive network measurements. We show, however, that the FB method is accurate only if the TCP transfer is window-limited to the point that it does not saturate the underlying path, and explain the main causes of the prediction errors. HB techniques predict the throughput of TCP flows from a time series of previous TCP throughput measurements on the same path, when such a history is available. We show that even simple HB predictors, such as Moving Average and Holt-Winters, using a history of few and sporadic samples, can be quite accurate. On the negative side, HB predictors are highly path-dependent. We explain the cause of such path dependencies based on two key factors: the load on the path and the degree of statistical multiplexing.

No abstract available.

The available bandwidth (avail-bw) of a network path is an important performance metric and its end-to-end estimation has recently received significant attention. Previous work focused on the estimation of the average avail-bw, ignoring the significant variability of this metric in different time scales. In this paper, we show how to estimate a given percentile of the avail-bw distribution at a user-specified time scale. If two estimated percentiles cover the bulk of the distribution (say 10% to 90%), the user can obtain a practical estimate for the avail-bw variation range. We present two estimation techniques. The first is iterative and non-parametric, meaning that it is more appropriate for very short time scales (typically less than 100ms), or in bottlenecks with limited flow multiplexing (where the avail-bw distribution may be non-Gaussian). The second technique is parametric, because it assumes that the avail-bw follows the Gaussian distribution, and it can produce an estimate faster because it is not iterative. The two techniques have been implemented in a measurement tool called Pathvar. Pathvar can track the avail-bw variation range within 10-20%, even under non-stationary conditions. We identify four factors that play a crucial role in the variation range of the avail-bw: traffic load, number of competing flows, rate of competing flows, and of course the measurement time scale. Finally, we present a new way to detect whether a probing rate is larger than the avail-bw, without relying on the fluid traffic assumption or on static thresholds.

Internet traffic exhibits multifaceted burstiness and correlation structure over a wide span of time scales. Previous work analyzed this structure in terms of heavy-tailed session characteristics, as well as TCP timeouts and congestion avoidance, in relatively long time scales. We focus on shorter scales, typically less than 100-1000 milliseconds. Our objective is to identify the actual mechanisms that are responsible for creating bursty traffic in those scales. We show that TCP self-clocking, joint with queueing in the network, can shape the packet interarrivals of a TCP connection in a two-level ON-OFF pattern. This structure creates strong correlations and burstiness in time scales that extend up to the Round-Trip Time (RTT) of the connection. This effect is more important for bulk transfers that have a large bandwidth-delay product relative to their window size. Also, the aggregation of many flows, without rescaling their packet interarrivals, does not converge to a Poisson stream, as one might expect from classical superposition results. Instead, the burstiness in those scales can be significantly reduced by TCP pacing. In particular, we focus on the importance of the minimum pacing timer, and show that a 10-millisecond timer would be too coarse for removing short-scale traffic burstiness, while a 1-millisecond timer would be sufficient to make the traffic almost as smooth as a Poisson stream in sub-RTT scales.

The area of available bandwidth (avail-bw) estimation has attracted significant interest recently, with several estimation techniques and tools developed during the last 2-3 years. Unfortunately, some key issues regarding the avail-bw definition, estimation, and validation remain vague or misinterpreted. In this note, we first review the previous work in the area and classify the existing techniques in two classes: direct probing and iterative probing. We then identify ten misconceptions, in the form of fallacies or pitfalls, that we consider as most important. Some misconceptions relate to basic statistics, such as the impact of the population variance on the sample mean, the variability of the avail-bw in different time scales, and the effect of the probing duration. Other misconceptions relate to the queueing model underlying these estimation techniques. For instance, ignoring that traffic burstiness or the presence of multiple bottlenecks can cause significant underestimation errors. Our objective is not to debunk previous work or to claim that some estimation techniques are better than others, but to clarify a number of important issues that cover the entire area of avail-bw estimation so that this important metric can be better understood and put in practical use.

The packet-pair technique aims to estimate the capacity of a path (bottleneck bandwidth) from the dispersion of two equal-sized probing packets sent back to back. It has been also argued that the dispersion of longer packet bursts (packet trains) can estimate the available bandwidth of a path. This paper examines such packet-pair and packet-train dispersion techniques in depth. We first demonstrate that, in general, packet-pair bandwidth measurements follow a multimodal distribution and explain the causes of multiple local modes. The path capacity is a local mode, often different than the global mode of this distribution. We illustrate the effects of network load, cross-traffic packet-size variability, and probing packet size on the bandwidth distribution of packet pairs. We then switch to the dispersion of long packet trains. The mean of the packet-train dispersion distribution corresponds to a bandwidth metric that we refer to as average dispersion rate (ADR). We show that the ADR is a lower bound of the capacity and an upper bound of the available bandwidth of a path. Putting all of the pieces together, we present a capacity-estimation methodology that has been implemented in a tool called pathrate. We report on our experiences with pathrate after having measured hundreds of Internet paths over the last three years.

By source-level IP packet burst, we mean several IP packets sent back-to-back from the source of a flow. We first identify several causes of source-level bursts, including TCP's slow start, idle restart, window advancement after loss recovery, and segmentation of application messages into multiple UDP packets. We then show that the presence of packet bursts in individual flows can have a major impact on aggregate traffic. In particular, such bursts create scaling in a range of timescales which corresponds to the burst duration. Uniform "spreading" of bursts in the time axis reduces the scaling exponent in short timescales (up to 100-200ms) to almost zero, meaning that the aggregate traffic becomes practically uncorrelated in that range. This result provides a plausible explanation for the scaling behavior of Internet traffic in short timescales. We also show that removing packet bursts from individual flows reduces significantly the tail of the aggregate marginal distribution, and it improves queueing performance, especially in moderate utilizations (50-85%).

Tools such as pathchar, clink, and pchar attempt to measure the capacity of every layer-3 (L3) hop in a network path. These tools use the same underlying measurement methodology, which we refer to as Variable Packet Size (VPS) probing. The key assumption in VPS is that each L3 hop along a path increases the delay of a packet by a "serialization latency", which is the ratio of the packet size over that hop's capacity. Unfortunately, the capacity estimates of VPS tools are sometimes wrong. In this paper, we investigate the source of these errors, and show that the presence of layer-2 (L2) store-and-forward devices, such as Ethernet switches, have a detrimental effect on the accuracy of VPS tools. Specifically, each L2 store-and-forward device introduces an additional serialization latency in a packet's delay, which results in consistent underestimation of that L3 hop's capacity. We analyze this negative effect, deriving the measured capacity of an L3 hop as a function of the L2 link capacities at that hop. Experimental results in local, campus, and ISP networks verify the model, illustrating that L2 devices should be expected in networks of diverse type and size. Finally, we characterize some other sources of error in VPS tools, such as queueing delays, limited clock resolution, variation in ICMP generation delays, and error propagation along the measured path.

The available bandwidth (avail-bw) in a network path is of major importance in congestion control, streaming applications, quality-of-service verification, server selection, and overlay networks. We describe an end-to-end methodology, called self-loading periodic streams (SLoPS), for measuring avail-bw. The basic idea in SLoPS is that the one-way delays of a periodic packet stream show an increasing trend when the stream's rate is higher than the avail-bw. We have implemented SLoPS in a tool called pathload. The accuracy of the tool has been evaluated with both simulations and experiments over real-world Internet paths. Pathload is nonintrusive, meaning that it does not cause significant increases in the network utilization, delays, or losses. We used pathload to evaluate the variability ("dynamics") of the avail-bw in Internet paths. The avail-bw becomes significantly more variable in heavily utilized paths, as well as in paths with limited capacity (probably due to a lower degree of statistical multiplexing). We finally examine the relation between avail-bw and TCP throughput. A persistent TCP connection can be used to measure roughly the avail-bw in a path, but TCP saturates the path and increases significantly the path delays and jitter.

No abstract available.

The available bandwidth (avail-bw) in a network path is of major importance in congestion control, streaming applications, QoS verification, server selection, and overlay networks. We describe an end-to-end methodology, called Self-Loading Periodic Streams (SLoPS), for measuring avail-bw. The basic idea in SLoPS is that the one-way delays of a periodic packet stream show an increasing trend when the stream's rate is higher than the avail-bw. We implemented SLoPS in a tool called pathload. The accuracy of the tool has been evaluated with both simulations and experiments over real-world Internet paths. Pathload is non-intrusive, meaning that it does not cause significant increases in the network utilization, delays, or losses. We used pathload to evaluate the variability ('dynamics') of the avail-bw in some paths that cross USA and Europe. The avail-bw becomes significantly more variable in heavily utilized paths, as well as in paths with limited capacity (probably due to a lower degree of statistical multiplexing). We finally examine the relation between avail-bw and TCP throughput. A persistent TCP connection can be used to roughly measure the avail-bw in a path, but TCP saturates the path, and increases significantly the path delays and jitter.

The proportional differentiation model provides the network operator with the 'tuning knobs' for adjusting the per-hop quality-of-service (QoS) ratios between classes, independent of the class loads. This paper applies the proportional model in the differentiation of queueing delays, and investigates appropriate packet scheduling mechanisms. Starting from the proportional delay differentiation (PDD) model, we derive the average queueing delay in each class, show the dynamics of the class delays under the PDD constraints, and state the conditions in which the PDD model is feasible. The feasibility model of the model can be determined from the average delays that result with the strict priorities scheduler. We then focus on scheduling mechanisms that can implement the PDD model, when it is feasible to do so. The proportional average delay (PAD) scheduler meets the PDD constraints, when they are feasible, but it exhibits a pathological behavior in short timescales. The waiting time priority (WTP) scheduler, on the other hand, approximates the PDD model closely, even in the short timescales of a few packet departures, but only in heavy load conditions. PAD and WTP serve as motivation for the third scheduler, called hybrid proportional delay (HPD). HPD approximates the PDD model closely, when the model is feasible, independent of the class load distribution. Also, HPD provides predictable delay differentiation even in short timescales.

The relative differentiation architecture does not require per-flow state at the network core or edges, nor admission control, but it can only provide higher classes with better service than lower classes. A central premise in this context is that users with absolute QoS requirements should search dynamically for an appropriate class. We investigate this dynamic class selection (DCS) framework, and illustrate that, under certain conditions, DCS-capable users can meet absolute QoS requirements, even though the network only offers relative differentiation. For a single link model, we can examine whether it is feasible to satisfy all users, and when this is the case, compute the minimum acceptable class selection for each user. Users converge in a distributed manner to this minimum acceptable class, if the DCS equilibrium is unique. However, suboptimal DCS equilibria may also exist. Simulations of a delay-based DCS algorithm show the relation between class differentiation and DCS, and demonstrate how to control the trade-off between the performance and cost of a flow.

The packet pair technique estimates the capacity of a path (bottleneck bandwidth) from the dispersion (spacing) experienced by two back-to-back packets. We demonstrate that the dispersion of packet pairs in loaded paths follows a multimodal distribution, and discuss the queueing effects that cause the multiple modes. We show that the path capacity is often not the global mode, and so it cannot be estimated using standard statistical procedures. The effect of the size of the probing packets is also investigated, showing that the conventional wisdom of using maximum sized packet pairs is not optimal. We then study the dispersion of long packet trains. Increasing the length of the packet train reduces the measurement variance, but the estimates converge to a value, referred to as the asymptotic dispersion rate (ADR), that is lower than the capacity. We derive the effect of the cross traffic in the dispersion of long packet trains, showing that the ADR is not the available bandwidth in the path, as was assumed in previous work. Putting all the pieces together, we present a capacity estimation methodology that has been implemented in a tool called pathrate.

Internet applications and users have very diverse service expectations, making the current same-service-to-all model inadequate and limiting. In the relative differentiated services approach, the network traffic is grouped in a small number of service classes which are ordered based on their packet forwarding quality, in terms of per-hop metrics for the queueing delays and packet losses. The users and applications, in this context, can adaptivelychoose the class that best meets their quality and pricing constraints, based on the assurance that higher classes will be better, or at least no worse, than lower classes. In this work, we propose the proportional differentiation model as a way to refine and quantify this basic premise of relative differentiated services. The proportional differentiation model aims to provide the network operator with the 'tuning knobs' for adjusting the quality spacing between classes, independent of the class loads; this cannot be achieved with other relative differentiation models, such as strict prioritization or capacity differentiation. We apply the proportional model on queueing-delay differentiation only, leaving the problem of coupled delay and loss differentiation for future work. We discuss the dynamics of the proportional delay differentiation model and state the conditions under which it is feasible. Then, we identify and evaluate (using simulations) two packet schedulers that approximate the proportional differentiation model in heavy-load conditions, even in short timescales. Finally, we demonstrate that such per-hop and class-based mechanisms can provide consistent end-to-end differentiation to individual flows from different classes, independently of the network path and flow characteristics.

No abstract available.