In 2016, Google proposed and deployed a new TCP variant called BBR. BBR represents a major departure from traditional congestion-window-based congestion control. Instead of using loss as a congestion signal, BBR uses estimates of the bandwidth and round-trip delays to regulate its sending rate. The last major study on the distribution of TCP variants on the Internet was done in 2011, so it is timely to conduct a new census given the recent developments around BBR. To this end, we designed and implemented Gordon, a tool that allows us to measure the exact congestion window (cwnd) corresponding to each successive RTT in the TCP connection response of a congestion control algorithm. To compare a measured flow to the known variants, we created a localized bottleneck where we can introduce a variety of network changes like loss events, bandwidth change, and increased delay, and normalize all measurements by RTT. An offline classifier is used to identify the TCP variant based on the cwnd trace over time. Our results suggest that CUBIC is currently the dominant TCP variant on the Internet, and it is deployed on about 36% of the websites in the Alexa Top 20,000 list. While BBR and its variant BBR G1.1 are currently in second place with a 22% share by website count, their present share of total Internet traffic volume is estimated to be larger than 40%. We also found that Akamai has deployed a unique loss-agnostic rate-based TCP variant on some 6% of the Alexa Top 20,000 websites and there are likely other undocumented variants. The traditional assumption that TCP variants "in the wild" will come from a small known set is not likely to be true anymore. We predict that some variant of BBR seems poised to replace CUBIC as the next dominant TCP variant on the Internet.
We consider the problem of forecasting fine-grained company financials, such as daily revenue, from two input types: noisy proxy signals a la alternative data (e.g. credit card transactions) and sparse ground-truth observations (e.g. quarterly earnings reports). We utilize a classical linear systems model to capture both the evolution of the hidden or latent state (e.g. daily revenue), as well as the proxy signal (e.g. credit cards transactions). The linear system model is particularly well suited here as data is extremely sparse (4 quarterly reports per year). In classical system identification, where the central theme is to learn parameters for such linear systems, unbiased and consistent estimation of parameters is not feasible: the likelihood is non-convex; and worse, the global optimum for maximum likelihood estimation is often non-unique. As the main contribution of this work, we provide a simple, consistent estimator of all parameters for the linear system model of interest; in addition the estimation is unbiased for some of the parameters. In effect, the additional sparse observations of aggregate hidden state (e.g. quarterly reports) enable system identification in our setup that is not feasible in general. For estimating and forecasting hidden state (actual earnings) using the noisy observations (daily credit card transactions), we utilize the learned linear model along with a natural adaptation of classical Kalman filtering (or Belief Propagation). This leads to optimal inference with respect to mean-squared error. Analytically, we argue that even though the underlying linear system may be "unstable," "uncontrollable," or "undetectable" in the classical setting, our setup and inference algorithm allow for estimation of hidden state with bounded error. Further, the estimation error of the algorithm monotonically decreases as the frequency of the sparse observations increases. This, seemingly intuitive insight contradicts the word on the Street. Finally, we utilize our framework to estimate quarterly earnings of 34 public companies using credit card transaction data. Our data-driven method convincingly outperforms the Wall Street consensus (analyst) estimates even though our method uses only credit card data as input, while the Wall Street consensus is based on various data sources including experts' input.
Heavy and highly dynamic traffic demands in today's data center networks (DCNs) pose great challenges to efficient traffic engineering. With gigabit bandwidth, wireless communication technologies, such as free space optics and 60GHz wireless, are promising to augment DCNs and enable efficient traffic engineering. Complementary to the emerging reconfigurable architectures, we aim to achieve efficient routing and effectively balance the load with the performance guarantee. We derive a general interference model and propose a decomposition technique with proven performance guarantee and solve the load balancing problem in reconfigurable DCNs. In addition, we propose two solutions, WiRo and OFS, to flexibly reconfigure network topology and enable hybrid-routing with paths consisting of both stable wired links and flexible wireless links with different methods. Our measurement-facilitated and trace-driven simulations demonstrate that our solutions outperform existing flow scheduling algorithms with the average throughput of large flows increased by up to 190% and the average completion time reduced by up to 72.6%. Meanwhile, the average completion time of small flows is reduced by up to 64.5%.
Classically, the design of multi-agent systems is approached using techniques from distributed optimization such as dual descent and consensus algorithms. Such algorithms depend on convergence to global consensus before any individual agent can determine its local action. This leads to challenges with respect to communication overhead and robustness, and improving algorithms with respect to these measures has been a focus of the community for decades.
This paper presents a new approach for multi-agent system design based on ideas from the emerging field of local computation algorithms. The framework we develop, LOcal Convex Optimization (LOCO), is the first local computation algorithm for convex optimization problems and can be applied in a wide-variety of settings. We demonstrate the generality of the framework via applications to Network Utility Maximization (NUM) and the distributed training of Support Vector Machines (SVMs), providing numerical results illustrating the improvement compared to classical distributed optimization approaches in each case.
Recently, researchers have started exploring the design of route protection schemes that ensure networks can sustain traffic demand without congestion under failures. Existing approaches focus on ensuring worst-case performance over simultaneous f-failure scenarios is acceptable. Unfortunately, even a single bad scenario may render the schemes unable to protect against any f-failure scenario. In this paper, we present Lancet, a system designed to handle most failures when not all can be tackled. Lancet comprises three components: (i) an algorithm to analyze which failure scenarios the network can intrinsically handle which provides a benchmark for any protection routing scheme, and guides the design of new schemes; (ii) an approach to efficiently design a protection schemes for more general failure sets than all f-failure scenarios; and (iii) techniques to determine which of combinatorially many scenarios to design for. Our evaluations with real topologies and validations on an emulation testbed show that Lancet outperforms a worst-case approach by protecting against many more scenarios, and can even match the scenarios that can be handled by optimal network response.
Volume-based network denial-of-service (DoS) attacks refer to a class of cyber attacks where an adversary seeks to block user traffic from service by sending adversarial traffic that reduces the available user capacity. In this paper, we explore the fundamental limits of volume-based network DoS attacks by studying the minimum required rate of adversarial traffic and investigating optimal attack strategies. We start our analysis with single-hop networks where user traffic is routed to servers following the Join-the-Shortest-Queue (JSQ) rule. Given the service rates of servers and arrival rates of user traffic, we first characterize the feasibility region of the attack and show that the attack is feasible if and only if the rate of the adversarial traffic lies in the region. We then design an attack strategy that is (i).optimal: it guarantees the success of the attack whenever the adversarial traffic rate lies in the feasibility region and (ii).oblivious: it does not rely on knowledge of service rates or user traffic rates. Finally, we extend our results on the feasibility region of the attack and the optimal attack strategy to multi-hop networks that employ Back-pressure (Max-Weight) routing. At a higher level, this paper addresses a class of dual problems of stochastic network stability, i.e., how to optimally de-stabilize a network.
Traffic measurement provides critical information for network management, resource allocation, traffic engineering, and attack detection. Most prior art has been geared towards specific application needs with specific performance objectives. To support diverse requirements with efficient and future-proof implementation, this paper takes a new approach to establish common frameworks, each for a family of traffic measurement solutions that share the same implementation structure, providing a high level of generality, for both size and spread measurements and for all flows. The designs support many options of performance-overhead tradeoff with as few as one memory update per packet and as little space as several bits per flow on average. Such a family-based approach will unify implementation by removing redundancy from different measurement tasks and support reconfigurability in a plug-n-play manner. We demonstrate the connection and difference in the design of these traffic measurement families and perform experimental comparisons on hardware/software platforms to find their tradeoff, which provide practical guidance for which solutions to use under given performance goals.
In the distributed graident coding problem, it has been established that, to exactly recover the gradient under s slow machines, the mmimum computation load (number of stored data partitions) of each worker is at least linear ($s+1$), which incurs a large overhead when s is large~\citetandon2017gradient. In this paper, we focus on approximate gradient coding that aims to recover the gradient with bounded error ε. Theoretically, our main contributions are three-fold: (i) we analyze the structure of optimal gradient codes, and derive the information-theoretical lower bound of minimum computation load: $O(łog(n)/łog(n/s))$ for ε = 0$ and $d\geq O(łog(1/ε)/łog(n/s))$ for ε>0$, where d is the computation load, and ε is the error in the gradient computation; (ii) we design two approximate gradient coding schemes that exactly match such lower bounds based on random edge removal process; (iii) we implement our schemes and demonstrate the advantage of the approaches over the current fastest gradient coding strategies. The proposed schemes provide order-wise improvement over the state of the art in terms of computation load, and are also optimal in terms of both computation load and latency.
Motivated by emerging need of learning algorithms for large scale networked and decentralized systems, we introduce a distributed version of the classical stochastic Multi-Arm Bandit (MAB) problem. Our setting consists of a large number of agents n that collaboratively and simultaneously solve the same instance of K armed MAB to minimize the average cumulative regret over all agents. The agents can communicate and collaborate among each other only through a pairwise asynchronous gossip based protocol that exchange a limited number of bits. In our model, agents at each point decide on (i) which arm to play, (ii) whether to, and if so (iii) what and whom to communicate with. Agents in our model are decentralized, namely their actions only depend on their observed history in the past. We develop a novel algorithm in which agents, whenever they choose, communicate only arm-ids and not samples, with another agent chosen uniformly and independently at random. The per-agent regret scaling achieved by our algorithm is $\BigO łeft( \fracłceil\fracK n \rceil+łog(n) Δ łog(T) + \fracłog^3(n) łog łog(n) Δ^2 \right) $. Furthermore, any agent in our algorithm communicates (arm-ids to an uniformly and independently chosen agent) only a total of Θ(łog(T))$ times over a time interval of T. We compare our results to two benchmarks - one where there is no communication among agents and one corresponding to complete interaction, where an agent has access to the entire system history of arms played and rewards obtained of all agents. We show both theoretically and empirically, that our algorithm experiences a significant reduction both in per-agent regret when compared to the case when agents do not collaborate and each agent is playing the standard MAB problem (where regret would scale linearly in K), and in communication complexity when compared to the full interaction setting which requires T communication attempts by an agent over T arm pulls. Our result thus demonstrates that even a minimal level of collaboration among the different agents enables a significant reduction in per-agent regret.
We determine information theoretic conditions under which it is possible to partially recover the alignment used to generate a pair of sparse, correlated Erdos-Renyi graphs. To prove our achievability result, we introduce the k-core alignment estimator. This estimator searches for an alignment in which the intersection of the correlated graphs using this alignment has a minimum degree of k. We prove a matching converse bound. As the number of vertices grows, recovery of the alignment for a fraction of the vertices tending to one is possible when the average degree of the intersection of the graph pair tends to infinity. It was previously known that exact alignment is possible when this average degree grows faster than the logarithm of the number of vertices.
We present an efficient algorithm for model-free episodic reinforcement learning on large (potentially continuous) state-action spaces. Our algorithm is based on a novel Q-learning policy with adaptive data-driven discretization. The central idea is to maintain a finer partition of the state-action space in regions which are frequently visited in historical trajectories, and have higher payoff estimates. We demonstrate how our adaptive partitions take advantage of the shape of the optimal Q-function and the joint space, without sacrificing the worst-case performance. In particular, we recover the regret guarantees of prior algorithms for continuous state-action spaces, which additionally require either an optimal discretization as input, and/or access to a simulation oracle. Moreover, experiments demonstrate how our algorithm automatically adapts to the underlying structure of the problem, resulting in much better performance compared both to heuristics and Q-learning with uniform discretization.
Inferring the quality of streaming video applications is important for Internet service providers, but the fact that most video streams are encrypted makes it difficult to do so. We develop models that infer quality metrics (\ie, startup delay and resolution) for encrypted streaming video services. Our paper builds on previous work, but extends it in several ways. First, the models work in deployment settings where the video sessions and segments must be identified from a mix of traffic and the time precision of the collected traffic statistics is more coarse (\eg, due to aggregation). Second, we develop a single composite model that works for a range of different services (\ie, Netflix, YouTube, Amazon, and Twitch), as opposed to just a single service. Third, unlike many previous models, our models perform predictions at finer granularity (\eg, the precise startup delay instead of just detecting short versus long delays) allowing to draw better conclusions on the ongoing streaming quality. Fourth, we demonstrate the models are practical through a 16-month deployment in 66 homes and provide new insights about the relationships between Internet "speed'' and the quality of the corresponding video streams, for a variety of services; we find that higher speeds provide only minimal improvements to startup delay and resolution.
Load balancing plays a crucial role in many large scale computer systems. Much prior work has focused on systems with First-Come-First-Served (FCFS) servers. However, servers in practical systems are more complicated. They serve multiple jobs at once, and their service rate can depend on the number of jobs in service. Motivated by this, we study load balancing for systems using Limited-Processor-Sharing (LPS). Our model has heterogeneous servers, meaning the service rate curve and multiprogramming level (limit on the number of jobs sharing the processor) differs between servers. We focus on a specific load balancing policy: Join-Below-Threshold (JBT), which associates a threshold with each server and, whenever possible, dispatches to a server which has fewer jobs than its threshold. Given this setup, we ask: how should we configure the system to optimize objectives such as mean response time? Configuring the system means choosing both a load balancing threshold and a multiprogramming level for each server. To make this question tractable, we study the many-server mean field regime. In this paper we provide a comprehensive study of JBT in the mean field regime. We begin by developing a mean field model for the case of exponentially distributed job sizes. The evolution of our model is described by a differential inclusion, which complicates its analysis. We prove that the sequence of stationary measures of the finite systems converges to the fixed point of the differential inclusion, provided a unique fixed point exists. We derive simple conditions on the service rate curves to guarantee the existence of a unique fixed point. We demonstrate that when these conditions are not satisfied, there may be multiple fixed points, meaning metastability may occur. Finally, we give a simple method for determining the optimal system configuration to minimize the mean response time and related metrics. While our theoretical results are proven for the special case of exponentially distributed job sizes, we provide evidence from simulation that the system becomes insensitive to the job size distribution in the mean field regime, suggesting our results are more generally applicable.
Large-scale machine learning and data mining applications require computer systems to perform massive matrix-vector and matrix-matrix multiplication operations that need to be parallelized across multiple nodes. The presence of straggling nodes -- computing nodes that unpredictably slowdown or fail -- is a major bottleneck in such distributed computations. Ideal load balancing strategies that dynamically allocate more tasks to faster nodes require knowledge or monitoring of node speeds as well as the ability to quickly move data. Recently proposed fixed-rate erasure coding strategies can handle unpredictable node slowdown, but they ignore partial work done by straggling nodes thus resulting in a lot of redundant computation. We propose a rateless fountain coding strategy that achieves the best of both worlds -- we prove that its latency is asymptotically equal to ideal load balancing, and it performs asymptotically zero redundant computations. Our idea is to create linear combinations of the m rows of the matrix and assign these encoded rows to different worker nodes. The original matrix-vector product can be decoded as soon as slightly more than m row-vector products are collectively finished by the nodes. We conduct experiments in three computing environments: local parallel computing, Amazon EC2, and Amazon Lambda, which show that rateless coding gives as much as 3x speed-up over uncoded schemes.
In this paper, we introduce theTheory of Bottleneck Ordering, a mathematical framework that reveals the bottleneck structure of data networks. This theoretical framework provides insights into the inherent topological properties of a network in at least three areas: (1) It identifies the regions of influence of each bottleneck; (2) it reveals the order in which bottlenecks (and flows traversing them) converge to their steady state transmission rates in distributed congestion control algorithms; and (3) it provides key insights into the design of optimized traffic engineering policies. We demonstrate the efficacy of the proposed theory in TCP congestion-controlled networks for two broad classes of algorithms: Congestion-based algorithms (TCP BBR) and loss-based additive-increase/multiplicative-decrease algorithms (TCP Cubic and Reno). Among other results, our network experiments show that: (1) Qualitatively, both classes of congestion control algorithms behave as predicted by the bottleneck structure of the network; (2) flows compete for bandwidth only with other flows operating at the same bottleneck level; (3) BBR flows achieve higher performance and fairness than Cubic and Reno flows due to their ability to operate at the right bottleneck level; (4) the bottleneck structure of a network is continuously changing and its levels can be folded due to variations in the flows' round trip times; and (5) against conventional wisdom, low-hitter flows can have a large impact to the overall performance of a network.
It has become increasingly difficult to understand the complex interactions between modern applications and main memory, composed of Dynamic Random Access Memory (DRAM) chips. Manufacturers are now selling and proposing many different types of DRAM, with each DRAM type catering to different needs (e.g., high throughput, low power, high memory density). At the same time, memory access patterns of prevalent and emerging applications are rapidly diverging, as these applications manipulate larger data sets in very different ways. As a result, the combined DRAM-workload behavior is often difficult to intuitively determine today, which can hinder memory optimizations in both hardware and software. In this work, we identify important families of workloads, as well as prevalent types of DRAM chips, and rigorously analyze the combined DRAM-workload behavior. To this end, we perform a comprehensive experimental study of the interaction between nine different DRAM types and 115 modern applications and multiprogrammed workloads. We draw 12 key observations from our characterization, enabled in part by our development of new metrics that take into account contention between memory requests due to hardware design. Notably, we find that (1) newer DRAM technologies such as DDR4 and HMC often do not outperform older technologies such as DDR3, due to higher access latencies and, also in the case of HMC, poor exploitation of locality; (2) there is no single memory type that can effectively cater to all of the components of a heterogeneous system (e.g., GDDR5 significantly outperforms other memories for multimedia acceleration, while HMC significantly outperforms other memories for network acceleration); and (3) there is still a strong need to lower DRAM latency, but unfortunately the current design trend of commodity DRAM is toward higher latencies to obtain other benefits. We hope that the trends we identify can drive optimizations in both hardware and software design. To aid further study, we open-source our extensively-modified simulator, as well as a benchmark suite containing our applications.
Modern network telemetry systems collect and analyze massive amounts of raw data in a space efficient manner. These require advanced capabilities such as drill down queries that allow iterative refinement of the search space. We present a first integral solution that (i) enables multiple measurement tasks inside the same data structure, (ii) supports specifying the time frame of interest as part of its queries, and (iii) is sketch-based and thus space efficient. Namely, our approach allows the user to define both the measurement task (e.g., heavy hitters, entropy estimation, count distinct, etc.) and the time frame of relevance (e.g., 5PM-6PM) at query time. Our approach provides accuracy guarantees and is the only space-efficient solution that offers such capabilities. Finally, we demonstrate how our system can be used for accurately pinpointing the start of a realistic DDoS attack.