The two largest U.S. wireless ISPs have recently moved towards usage-based pricing to better manage the growing demand on their networks. Yet still requires over-provision capacity for at peak times of day. Time-dependent (TDP) addresses this problem by considering when a user consumes data, in addition how much is used. We present architecture, implementation, and trial an end-to-end TDP system called TUBE. TUBE creates price-based feedback control loop between ISP its end users. On side,...

Deep neural networks (DNN) are the de-facto solution behind many intelligent applications of today, ranging from machine translation to autonomous driving. DNNs accurate but resource-intensive, especially for embedded devices such as mobile phones and smart objects in Internet Things. To overcome related resource constraints, DNN inference is generally offloaded edge or cloud. This accomplished by partitioning distributing computations at two different ends. However, most existing solutions...

Smart grids are capable of two-way communication between individual user devices and the electricity provider, enabling providers to create a control-feedback loop using time-dependent pricing. By charging users more in peak less off-peak hours, provider can induce shift their consumption periods, thus relieving stress on power grid cost incurred from large loads. We formulate provider's minimization problem setting these prices by considering consumers' device-specific scheduling...

Amazon's Elastic Compute Cloud (EC2) uses auction-based spot pricing to sell spare capacity, allowing users bid for cloud resources at a highly reduced rate. Amazon sets the price dynamically and accepts user bids above this price. Jobs with lower (including those already running) are interrupted must wait before resuming. Spot thus raises two basic questions: how might provider set price, what prices should bid? Computing users' bidding strategies is particularly challenging: higher reduce...

Quantifying the notion of fairness is under-explored when users request different ratios multiple distinct resource types. A typical example datacenters processing jobs with heterogeneous requirements on CPU, memory, etc. generalization max-min to resources was recently proposed in [1], but may suffer from significant loss efficiency. This paper develops a unifying framework addressing this fairness-efficiency tradeoff We develop two families functions which provide tradeoffs, characterize...

The tremendous growth in demand for broadband data is forcing ISPs to use pricing as a congestion management tool. This changing landscape of Internet access evidenced by the elimination flat rate plans favor usage-based major wired and wireless operators US Europe. But simple fees suffer from problem imposing costs on all users, irrespective network level at given time. To effectively reduce congestion, appropriate incentives must be provided users who are willing time-shift their peak...

Quantifying the notion of fairness is underexplored when there are multiple types resources and users request different ratios resources. A typical example data centers processing jobs with heterogeneous resource requirements on CPU, memory, network bandwidth, etc. In such cases, a tradeoff arises between equitability, or “fairness,” efficiency. This paper develops unifying framework addressing fairness-efficiency in light We develop two families functions that provide tradeoffs,...

We present a novel method for predicting the evolution of student's grade in massive open online courses (MOOCs). Performance prediction is particularly challenging MOOC settings due to per-student assessment response sparsity and need personalized models. Our overcomes these challenges by incorporating another, richer form data collected from each student-lecture video-watching clickstreams-into machine learning feature set, using that train time series neural network learns both prior...

In January 2014, AT&T introduced sponsored data to the U.S. mobile market, allowing content providers (CPs) subsidize users' cost of data. As gains traction in industry, it is important understand its implications. This work considers CPs' choice how much sponsor and implications for users, CPs, ISPs (Internet service providers). We first formulate a model user, CP, ISP interaction heterogeneous users CPs derive their optimal behaviors. then show that these behaviors can reverse our...

Vehicular crowd sensing systems are designed to achieve large spatio-temporal coverage with low-cost in deployment and maintenance. For example, taxi platforms can be utilized for city-wide air quality. However, the goals of vehicle agents often inconsistent goal crowdsourcer. Vehicle like taxis prioritize searching passenger ride requests (defined as task requests), which leads them gather busy regions. In contrast, need sample data over entire city a desired distribution (e.g., Uniform...

Modern vehicle fleets, e.g., for ridesharing platforms and taxi companies, can reduce passengers' waiting times by proactively dispatching vehicles to locations where pickup requests are anticipated in the future. Yet it is unclear how best do this: optimal requires optimizing over several sources of uncertainty, including vehicles' travel their dispatched locations, as well coordinating between so that they not attempt pick up same passenger. While prior works have developed models this...

Vehicular mobile crowdsensing (MCS) enables many smart city applications. Ridesharing vehicle fleets provide promising solutions to MCS due the advantages of low cost, easy maintenance, high mobility, and long operational time. However, as nondedicated sensing platforms, first priorities these vehicles are delivering passengers, which may lead poor coverage quality. Therefore, help derive good (large balanced) quality, an actuation system is required dispatch with a limited amount monetary...

Traditionally, clustered federated learning groups clients with the same data distribution into a cluster, so that every client is uniquely associated one and helps train model for this distribution. We relax hard association assumption to soft learning, which allows local dataset follow mixture of multiple source distributions. propose FedSoft, trains both locally personalized models high-quality cluster in setting. FedSoft limits workload by using proximal updates require completion only...

The two largest U.S. wireless ISPs have recently moved towards usage-based pricing to better manage the growing demand on their networks. Yet still requires over-provision capacity for at peak times of day. Time-dependent (TDP) addresses this problem by considering when a user consumes data, in addition how much is used. We present architecture, implementation, and trial an end-to-end TDP system called TUBE. TUBE creates price-based feedback control loop between ISP its end users. On side,...

Fog computing promises to enable machine learning tasks scale large amounts of data by distributing processing across connected devices. Two key challenges achieving this are (i) heterogeneity in devices' compute resources and (ii) topology constraints on which devices can communicate. We the first address these developing a network-aware distributed optimization methodology where process for task locally send their learnt parameters server aggregation at certain time intervals. Unlike...

Fog computing promises to enable machine learning tasks scale large amounts of data by distributing processing across connected devices. Two key challenges achieving this goal are (i) heterogeneity in devices' compute resources and (ii) topology constraints on which devices communicate with each other. We address these developing a novel network-aware distributed methodology where optimally share local send their learnt parameters server for periodic aggregation. Unlike traditional federated...

Federated Learning trains machine learning models on distributed devices by aggregating local model updates instead of data. However, privacy concerns arise as the aggregated server may reveal sensitive personal information inversion attacks. Privacy-preserving methods, such homomorphic encryption (HE), then become necessary for FL training. Despite HE's advantages, its applications suffer from impractical overheads, especially foundation models. In this paper, we present FedML-HE, first...

This paper introduces FedSecurity, an end-to-end benchmark that serves as a supplementary component of the FedML library for simulating adversarial attacks and corresponding defense mechanisms in Federated Learning (FL). FedSecurity eliminates need implementing fundamental FL procedures, e.g., training data loading, from scratch, thus enables users to focus on developing their own attack strategies. It contains two key components, including FedAttacker conducts variety during training,...

Charging different prices for Internet access at times induces users to spread out their bandwidth consumption across of the day. Potential impact on ISP revenue, congestion management, and consumer behavior can be significant, yet some fundamental questions remain: is it feasible operate time dependent pricing how much benefit bring? We develop an efficient way compute cost-minimizing time-dependent service provider (ISP), using both a static session-level model dynamic session with...

Amazon's Elastic Compute Cloud (EC2) uses auction-based spot pricing to sell spare capacity, allowing users bid for cloud resources at a highly reduced rate. Amazon sets the price dynamically and accepts user bids above this price. Jobs with lower (including those already running) are interrupted must wait before resuming. Spot thus raises two basic questions: how might provider set price, what prices should bid? Computing users' bidding strategies is particularly challenging: higher reduce...

In networked systems, initial failures at only a small part of the network may trigger sequential failure process called cascading failures, which eventually lead to breakdown entire system. This vulnerability is further exacerbated in modern systems that consist <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">multiple</i> networks, each carrying flow or load, are xmlns:xlink="http://www.w3.org/1999/xlink">interdependent</i> ; e.g.,...

Economic incentives that alleviate congestion for Internet customers can also improve business performance network operators.

The growing volume of mobile data traffic has led many Internet service providers (ISPs) to cap their users' monthly usage, with overage fees for exceeding caps. In this work, we examine a secondary market in which users can buy and sell leftover caps from each other. China Mobile Hong Kong recently introduced such market. While similar an auction that submit bids data, it differs traditional double auctions the ISP serves as middleman between buyers sellers. We derive optimal prices amount...

Cloud service providers (CSPs) often face highly dynamic user demands for their resources, which can make it difficult them to maintain consistent quality-of-service. Some CSPs try stabilize by offering sustained-use discounts jobs that consume more instance-hours per month. These present an opportunity users pool usage together into a single ``job.'' In this paper, we examine the viability of middleman, cloud virtual provider (CVSP), rents resources from CSP and then resells users. We show...