Search for: All records

Earth observation Low Earth Orbit (LEO) satellites collect enormous amounts of data that needs to be transferred first to ground stations and then to the cloud, for storage and processing. Satellites today transmit data greedily to ground stations, with full utilization of bandwidth during each contact period. We show that due to the layout of ground stations and orbital characteristics, this approach overloads some ground stations and underloads others, leading to lost throughput and large end-to-end latency for images. We present a new end-to-end scheduler system called Umbra, which plans transfers from large satellite constellations through ground stations to the cloud, by accounting for both spatial and temporal factors, i.e., orbital dynamics, bandwidth constraints, and queue sizes. At the heart of Umbra is a new class of scheduling algorithms called withhold scheduling, wherein the sender (i.e., satellite) selectively under-utilizes some links to ground stations. We show that Umbra’s counter-intuitive approach increases throughput by 13-31% & reduces P90 latency by 3-6 X. 

Distributed machine learning is primarily motivated by the promise of increased computation power for accelerating training and mitigating privacy concerns. Unlike machine learning on a single device, distributed machine learning requires collaboration and communication among the devices. This creates several new challenges: (1) the heavy communication overhead can be a bottleneck that slows down the training, and (2) the unreliable communication and weaker control over the remote entities make the distributed system vulnerable to systematic failures and malicious attacks. This paper presents a variant of stochastic gradient descent (SGD) with improved communication efficiency and security in distributed environments. Our contributions include (1) a new technique called error reset to adapt both infrequent synchronization and message compression for communication reduction in both synchronous and asynchronous training, (2) new score-based approaches for validating the updates, and (3) integration with both error reset and score-based validation. The proposed system provides communication reduction, both synchronous and asynchronous training, Byzantine tolerance, and local privacy preservation. We evaluate our techniques both theoretically and empirically. 

With the increasing adoption of smart home devices, users rely on device automation to control their homes. This automation commonly comes in the form of smart home routines, an abstraction available via major vendors. Yet, questions remain about how a system should best handle conflicts in which different routines access the same devices simultaneously. In particular---among the myriad ways a smart home system could handle conflicts, which of them are currently utilized by existing systems, and which ones result in the highest user satisfaction? We investigate the first question via a survey of existing literature and find a set of conditions, modifications, and system strategies related to handling conflicts. We answer the second question via a scenario-based Mechanical-Turk survey of users interested in owning smart home devices and current smart home device owners (N=197). We find that: (i) there is no context-agnostic strategy that always results in high user satisfaction, and (ii) users' personal values frequently form the basis for shaping their expectations of how routines should execute. 

Smart environments (homes, factories, hospitals, buildings) contain an increasing number of IoT devices, making them complex to manage. Today, in smart homes when users or triggers initiate routines (i.e., a sequence of commands), concurrent routines and device failures can cause incongruent outcomes. We describe SafeHome, a system that provides notions of atomicity and serial equivalence for smart homes. Due to the human-facing nature of smart homes, SafeHome offers a spectrum of visibility models which trade off between responsiveness vs. isolation of the smart home. We implemented SafeHome and performed workload-driven experiments. We find that a weak visibility model, called eventual visibility, is almost as fast as today's status quo (up to 23% slower) and yet guarantees serially-equivalent end states. 

Efficient and correct operation of an IoT network requires the presence of a failure detector and membership protocol amongst the IoT nodes. This paper presents a new failure de- tector for IoT settings where nodes are connected via a wire- less ad-hoc network. This failure detector, which we name Medley, is fully decentralized, allows IoT nodes to maintain a local membership list of other alive nodes, detects failures quickly (and updates the membership list), and incurs low communication overhead in the underlying ad-hoc network. In order to minimize detection time and communication, we adapt a failure detector originally proposed for datacenters (SWIM), for the IoT environment. In Medley each node picks a medley of ping targets in a randomized and skewed manner, preferring nearer nodes. Via analysis and NS-3 simulation we show the right mix of pinging probabilities that simulta- neously optimize detection time and communication traffic. We have also implemented Medley for Raspberry Pis, and present deployment results.