INFOCOM 2019 Papers

The need to scale up machine learning, in the presence of a rapid growth of data both in volume and in variety, has sparked broad interests to develop distributed machine learning systems, typically based on parameter servers. However, since these systems are based on a dedicated cluster of physical or virtual machines, they have posed non-trivial cluster management overhead to machine learning practitioners and data scientists. In addition, there exists an inherent mismatch between the dynamically varying resource demands during a model training job and the inflexible resource provisioning model of current cluster-based systems.In this paper, we propose SIREN, an asynchronous distributed machine learning framework based on the emerging serverless architecture, with which stateless functions can be executed in the cloud without the complexity of building and maintaining virtual machine infrastructures. With SIREN, we are able to achieve a higher level of parallelism and elasticity by using a swarm of stateless functions, each working on a different batch of data, while greatly reducing system configuration overhead. Furthermore, we propose a scheduler based on Deep Reinforcement Learning to dynamically control the number and memory size of the stateless functions that should be used in each training epoch. The scheduler learns from the training process itself, in pursuit for the minimum possible training time given a cost. With our real-world prototype implementation on AWS Lambda, extensive experimental results have shown that SIREN can reduce model training time by up to 44%, as compared to traditional machine learning training benchmarks on AWS EC2 at the same cost.

Mobile users and wireless devices are now the sources of a large volume of data. In such data-intensive mobile and wireless computing systems, delay-tolerant network (DTN) routing plays a critical role in data routing, dissemination, and collection. In this paper, we first introduce a new routing problem in DTNs - data-intensive routing - where data transmitted from one node to another is very large with respect to the size of data which can be transmitted in a single contact and available buffer size at relay nodes. In the proposed opportunistic path model, the contact frequency, contact duration, and buffer constraint are all integrated into a single routing metric. Then, we design the data-intensive routing (DIR) protocol where the path with the highest bottleneck link capacity is defined as the path weight. In addition, we propose the advanced DIR (A-DIR) protocol which focuses on the probability that the last message block will be delivered to its destination within the time constraint. Both the DIR and A-DIR protocols forward messages to better relays or to their destinations based on a greedy strategy with the proposed path metric. Simulations using real mobility traces demonstrate that the proposed DIR and A-DIR protocols achieve their design goals.

Network Function Virtualization (NFV) has the potential to offer service delivery flexibility and reduce overall costs by running service function chains (SFCs) on commodity servers with many cores. Existing solutions for placing SFCs in one server treat all CPU cores as equal and allocate isolated CPU cores to different network functions (NFs). However, advanced servers often adopt Non-Uniform Memory Access (NUMA) architecture to improve the scalability of many-core systems. CPU cores are grouped into nodes, incurring performance bottleneck due to cross-node memory access and intra-node resource contention. Our evaluation shows that randomly selecting cores to place NFs in an SFC could suffer from 39.2% lower throughput comparing to an optimal placement solution. In this paper, we propose Octans, an NFV orchestrator to achieve maximum aggregate throughput of all SFCs in many-core systems. Octans first formulates the optimization problem as a Non-Linear Integer Programming (NLIP) model. Then we identify the key factor for problem solving as evaluating the throughput drop of an NF caused by other NFs in the same SFC or different SFCs, i.e. performance drop index, and propose a formal and precise prediction model based on system level performance metrics. Finally, we propose an efficient heuristic algorithm to quickly find near-optimal placement solutions. We have implemented a prototype of Octans. Extensive evaluation shows that Octans significantly improves the aggregate throughput comparing to two state-of the-art placement mechanisms by 26.7%~51.8%, with very low prediction errors of SFC performance (an average deviation of 2.6%). Moreover, Octans could quickly find a near-optimal placement solution with tiny optimality gap (1.2%~3.5%).

Network utility maximization (NUM) is a general framework for optimally allocating constrained resources in many networked applications. When agents have asymmetric and private information, a fundamental economic challenge is how to solve the NUM Problem considering the self-interests of strategic agents. Many previous related works have proposed economic mechanisms that can cope with agents' private utilities. However, the related literature largely neglected the issue of information asymmetries regarding constraints, and limited closely related studies provided solutions only applicable to specific application scenarios. To tackle this issue, we propose the DeNUM Mechanism, the first mechanism for solving a general class of decomposable NUM Problems considering both private utility and constraint information. The key idea is to decentralize the decision process to agents, who will make resource allocation decisions without the need of revealing private information to others. We further show that the DeNUM mechanism yields the network-utility maximizing solution at an equilibrium, and achieves other desirable economic properties (such as individual rationality and budget balance). However, the corresponding equilibrium solution concept, the generalized Nash equilibrium (GNE), makes it difficult to achieve through a distributed algorithm. To address this issue, we further establish the connection between the structure of GNE and that of the primal-dual solution to a reformulated NUM problem, based on which we present the convergent DeNUM Algorithm that is provably convergent. Finally, as a case study, we apply the DeNUM Mechanism to solving the NUM problem for a user-provided network, and show that the DeNUM algorithm improves the network utility by 17% compared to a non-cooperation benchmark.

In this paper we propose an online algorithm for clustering channel-states and learning the associated achievable multiuser rates. Our motivation stems from the complexity of multiuser scheduling. For instance, MU-MIMO scheduling involves the selection of a user subset and associated rate selection each time-slot for varying channel states (the vector of quantized channels matrices for each of the users) - a complex integer optimization problem that is different for each channel state. Instead, our algorithm clusters the collection of channel states to a much lower dimension, and for each cluster provides achievable multiuser capacity trade-offs, which can be used for user and rate selection. Our algorithm uses a bandit approach, where it learns both the unknown partitions of the channel-state space (channel-state clustering) as well as the capacity region for each cluster along a pre-specified set of directions, by observing the success/failure of the scheduling decisions (e.g. through packet loss). We propose an epoch-greedy learning algorithm that achieves a sub-linear regret, given access to a class of classifying functions over the channel-state space. Finally, we empirically validate the performance of our algorithm through simulations.

Ambient environment information, including reflectors' location, dimension and reflectivity, is a key input to many millimeter-wave (mmWave) networking and sensing applications. It has found versatile applications in optimizing network coverage and robustness, enhancing mobile link performance, and enabling high-accuracy indoor localization and navigation. Recent approaches of deriving mmWave environment information require heavy infrastructure support or non-trivial human labor, and rely on costly software defined radios, which prevent their usage in practice. In this work, we design and implement mmRanger, a system can automatically sense environment without any infrastructure support. mmRanger equips a pair of low-cost off-the-shelf mmWave radios in a commodity robot, which constantly samples the ambient environment by exchanging a series of mmWave signals while it moves and rotates. mmRanger then re-engineers the time-domain signal series to derive the spatial-domain environment structure, through novel reflection path extraction and reflector mapping algorithms. Our experiments verify that mmRanger can accurately sense a given environment with minimal overhead, and the learned information can bring 1.6× and 2.1× performance gain, in terms of network coverage and mobile link throughput, respectively, over empirical approaches in mmWave networks.

Post-copy is one of the two key techniques (besides pre-copy) for live migration of virtual machines in data centers. Post-copy provides deterministic total migration time and low downtime for write-intensive VMs. However, if post-copy migration fails for any reason, the migrating VM is lost because the VM's latest consistent state is split between the source and destination nodes during migration. In this paper, we present PostCopyFT, a new approach to recover a VM after a destination or network failure during post-copy live migration using an efficient reverse incremental checkpointing mechanism. We have implemented and evaluated our approach in the KVM/QEMU platform. Our experimental results show that the total migration time of post-copy remains unchanged while maintaining low failover time, downtime, and application performance overhead.

Motivated by the fair rate allocation in a multiaccess Gaussian channel, this paper studies the problem of fair rate allocation over a generalized symmetric polymatroid with box constraints. The best-known algorithm for this problem has time complexity O(n5lnO(1)n). In this paper, we present a divide-and-conquer algorithm for this problem with quadratic running time. It is an implementation of a refined decomposing method for the more general separate concave maximization over a polymatroid with box constraints. A key ingredient of the algorithm is a linear-time algorithm for a generalized knapsack problem.

We consider the wireless scheduling problem of jointly activating/de-activating base-stations and (opportunistically) scheduling from among the active base stations. Such systems are of increasing relevance in emerging wireless networks with dense overlapping coverage, where it suffices for only a (time-varying) subset of the base-stations to be active at any given time to satisfy traffic demands. In addition to queue stability (to ensure that traffic demands are met), we focus on optimizing for costs arising due to activating base-stations (switching base-station state between active/inactive), and maintaining activation (these costs arising due to energy consumption).We propose two algorithms-LASS-Static and LASS-Dynamic (LASS: Learning Aided Switching and Scheduling), both of which are explore-exploit policies for base-station switching and channel scheduling. In our setting, the switching action consists of two key decisions: when to switch, and what base-station activation state to switch to. Both LASS-Static and LASS-Dynamic determine the resulting switching state (i.e. `what to switch to as well as the schedule using current queue-lengths and (estimated) channel states. The crucial difference is in `when to switch'-LASS-Static determines these statically (motivated by an epsilon-greedy bandit approach), whereas LASS-Dynamic does so using current queue-lengths (thus correlating switching times, switching states and schedules). For either algorithm, existing Lyapunov-based techniques fail to establish stability, as the switching state dynamics correlate the base-station activation decisions with the channel evolution over time. Using novel drift based techniques, in this paper we derive stability, and provide explicit bounds on the expected cost and queue lengths for both algorithms. Furthermore, we show that adaptively selecting switching times in LASS-Dynamic results in an improved upper-tail of queue lengths compared to LASS-Static.

With the ever-increasing complexity of networks, fine-grained network monitoring enables better network reliability and timely feedback control. The In-band Network Telemetry (INT) allows cost-effective network monitoring by encapsulating device-internal states into probe packets. However, INT only specifies an underlying device-level primitive while how to achieve network-wide traffic monitoring remains undefined. In this work, we propose INT-path, a network-wide telemetry framework, by decoupling the system into a routing mechanism and a routing path generation policy. Specifically, we embed source routing into INT probes to allow specifying the route the probe packet takes through the network. Above the mechanism, we develop an Euler trail-based path planning policy to generate non-overlapped INT paths that cover the entire network with a minimum path number. Besides, an exhaustive analysis of algorithm's run-time complexity is also provided. INT-path can “encode” the network-wide traffic status into a series of “bitmap images”, transforming network troubleshooting into pattern recognition problems. INT-path is very suitable for deployment in data center networks thanks to their symmetric network topologies.

In the framework of queuing theory with multiclass jobs, trunk-reservation is an admission control technique to handle job class priority in an online fashion, and serve as many high priority jobs as possible. This is achieved by rejecting jobs with sufficiently low priority when the buffer space becomes a scarce resource, i.e., when the queue may soon overflow. Mathematically, the objective is to maximize the long-term reward of the admitted jobs, where each job is assigned a reward which is monotonic with respect to the priority class it belongs to. In this paper we study online learning of optimal trunk-reservation policies when the system parameters, i.e., class arrival rates and service time, are unknown. Starting from a Markov Decision Process (MDP) formulation, we leverage the stairway structure of the optimal policy to define Integer Gradient Ascent (IGA), a reinforcement learning (RL) algorithm based on policy-gradient methods, specifically tailored to the mathematical properties of the problem at hand. We provide theoretical results on the convergence properties of IGA. Extensive numerical experiments characterize its behavior and confirm that it outperforms standard RL techniques in terms of convergence rate.

In this paper we consider the task of detecting abnormal communication volume occurring at node-level in communication networks. The signal of the communication activity is modeled by means of a clique stream: each occurring communication event is instantaneous, and activates an undirected subgraph spanning over a set of equally participating nodes. We present a probabilistic framework to model and assess the communication volume observed at any single node. Specifically, we employ non-parametric regression to learn the probability that a node takes part in a certain event knowing the set of other nodes that are involved. On the top of that, we present a concentration inequality around the estimated volume of events in which a node could participate, which in turn allows us to build an efficient and interpretable anomaly scoring function. Finally, the superior performance of the proposed approach is empirically demonstrated in real-world sensor network data, as well as using synthetic communication activity that is in accordance with that latter setting.

Many learning problems are formulated as minimization of some loss function on a training set of examples. Distributed gradient methods on a cluster are often used for this purpose. In this paper, we study how the variability of task execution times at cluster nodes affects the system throughput. In particular, a simple but accurate model allows us to quantity how the time to solve the minimization problem depends on the network of information exchanges among the nodes. Interestingly, we show that, even when communication overhead may be neglected, the clique is not necessarily the most effective topology, as commonly assumed in previous works.

Accountability and privacy are considered valuable but conflicting properties in the Internet, which at present does not provide native support for either. Past efforts to balance accountability and privacy in the Internet have unsatisfactory deployability due to the introduction of new communication identifiers, and because of large-scale modifications to fully deployed infrastructures and protocols. The IPv6 is being deployed around the world and this trend will accelerate. In this paper, we propose a private and accountable proposal based on IPv6 called PAVI that seeks to bootstrap accountability and privacy to the IPv6 Internet without introducing new communication identifiers and large-scale modifications to the deployed base. A dedicated quantitative analysis shows that the proposed PAVI achieves satisfactory levels of accountability and privacy. The results of evaluation of a PAVI prototype show that it incurs little performance overhead, and is widely deployable.

The recent proliferation of human-carried mobile devices has given rise to mobile crowd sensing (MCS) systems, where a myriad of data requesters outsource their sensing tasks to a crowd of workers via a cloud-based platform. In order to incentivize participation, requesters typically compensate workers with specific amount of payments. Clearly, setting an appropriate task price is critical for a requester to attract enough worker participation without unnecessary expenses. Therefore, we investigate the problem of task pricing in MCS systems with multi-requester price competition, and also dynamically arriving workers. Task pricing in such scenario is challenging, because of each requester's incomplete information about the others, uncertainty of future information, etc. So as to address these challenges, we use Markov game to model requesters' competitive task pricing, and Markov correlated equilibrium (MCE) as the solution concept. We propose that the platform uses the social cost-minimizing MCE to coordinate requesters' prices, which is self-enforcing, and optimizes the system-wide objective of social cost. Technically, we propose a computationally efficient algorithm to compute an approximately optimal MCE. Furthermore, through extensive performance evaluation, we show numerically that our algorithm yields close-to-minimum social cost in very short running time.

To deal with the increasingly larger input data and model sizes, it has become necessary to scale the training of machine learning models to multiple nodes, even a server cluster, which we call distributed machine learning, or DML. However, DML utilizes more computation power at the cost of high communication overhead, which may limit the overall performance in turn. In this paper, we study the impact of network topology on the DML performance both in theory and in practice. We compare two representative network topologies, namely, Fat-Tree which is widely-used in modern data centers, and BCube, which is a low-cost and server-centric network topology, both running on top of RDMA. The results show that Fat-Tree not only has theoretically higher global synchronization time (GST) than BCube, but its practical GST (by NS-3 based simulation) is also considerably larger than the theoretical one. By analyzing the large-scale simulation traces, we find that the root cause for the gap in Fat-Tree comes from the load imbalance among the multiple parallel paths as well as the inevitable PFC frames, both of which do not appear in BCube. For a cluster of around 250 servers, BCube achieves 53%\sim 70% lower GST than Fat-Tree from the simulation. As a result, we suggest using server-centric network topology such as BCube, instead of the common Fat-Tree network, to build a special-purpose DML cluster, due to its parallel synchronization, RDMA friendliness, natural load balance, as well as low economical cost.

The multi-armed bandit (MAB) model has been widely adopted for studying many practical optimization problems (network resource allocation, ad placement, crowdsourcing, etc.) with unknown parameters. The goal of the player (i.e., the decision maker) here is to maximize the cumulative reward in the face of uncertainty. However, the basic MAB model neglects several important factors of the system in many realworld applications, where multiple arms (i.e., actions) can be simultaneously played and an arm could sometimes be “sleeping” (i.e., unavailable). Besides reward maximization, ensuring fairness is also a key design concern in practice. To that end, we propose a new Combinatorial Sleeping MAB model with Fairness constraints, called CSMAB-F, aiming to address the aforementioned crucial modeling issues. The objective is now to maximize the reward while satisfying the fairness requirement of a minimum selection fraction for each individual arm. To tackle this new problem, we extend an online learning algorithm, called Upper Confidence Bound (UCB), to deal with a critical tradeoff between exploitation and exploration and employ the virtual queue technique to properly handle the fairness constraints. By carefully integrating these two techniques, we develop a new algorithm, called Learning with Fairness Guarantee (LFG), for the CSMAB-F problem. Further, we rigorously prove that not only LFG is feasibility-optimal, but it also has a time-average regret upper bounded by N/2η + β1√mNT log T +β2N/T, where N is the total T number of arms, m is the maximum number of arms that can be simultaneously played, T is the time horizon, β1and β2are constants, and η is a design parameter that we can tune. Finally, we perform extensive simulations to corroborate the effectiveness of the proposed algorithm. Interestingly, the simulation results reveal an important tradeoff between the regret and the speed of convergence to a point satisfying the fairness constraints.

This paper presents a novel network architecture, termed neXt-generation CHAllenged NeTwork (X-CHANT), for improving connectivity in rural environments. The underlying idea is to deploy diverse Dynamic Spectrum Access (d-DSA) radio devices on the public transportation vehicles, such as buses. This results in a d-DSA enabled delay-tolerant network in which the devices can operate in various (un)licensed bands (e.g., TV, LTE, ISM, CBRS), if available. Given the lack of research in efficient routing for such time-varying d-DSA enabled networks, we propose a novel diverse DSA aware routing (dDSAaR) protocol that jointly exploits various (un)licensed bands besides the time-varying yet sufficiently predictable mobility of public transportation vehicles. We compare X-CHANT, utilizing dDSAaR, to the conventional non-DSA/DSA architectures, utilizing a standard (single band) routing protocol (e.g., Epidemic). We use real bus mobility traces collected at the University of Massachusetts, Amherst campus. Results show that X-CHANT achieves better message delivery, negligible message overhead, and better energy expenditure, at the expense of a slight increase in delay. Never-theless, the delay improves with higher predictable mobility.

When facing fluctuating workloads, can the performance of matching algorithms in a content-based publish/subscribe system be adjusted to adapt to the workloads? In this paper, we explore the idea of endowing matching algorithms with adaptability. The prerequisite for adaptability is to enable the matching algorithm to possess the ability to dynamically and quantitatively adjust its performance. We propose PSAM, a Predicate-Skipping Adjustment Mechanism that realizes dynamic performance adjustment by smoothly switching between exact matching and approximate matching, following the strategy of trading off matching precision in favor of matching speed. The PSAM mechanism is integrated into an existing matching algorithm, resulting in a performance-adjustable matching algorithm called Ada-Rein. To collaborate with Ada-Rein, we design PADA, a Performance Adjustment Decision Algorithm that is able to make proper performance adjustment plans in the presence of fluctuating workloads. The effectiveness of Ada-Rein and PADA is evaluated through a series of experiments based on both synthetic data and real-world stock traces. Experiment results show that adjusting the performance of Ada-Rein at the price of a small false positive rate, less than 0.1%, can shorten event latency by almost 2.1 times, which well demonstrates the feasibility of our exploratory idea.