All sessions will be held in the Essex Ballroom Center unless otherwise noted.
Papers are available for download below to registered attendees now and to everyone beginning Tuesday, July 10, 2018. Paper abstracts are available to everyone now. Copyright to the individual works is retained by the author[s].
Downloads for Registered Attendees
(Sign in to your USENIX account to download these files.)
Tuesday, July 10, 2018
7:30 am–8:00 am
Essex Ballroom Foyer
8:00 am–8:15 am
Program Co-Chairs: Irfan Ahmad, CachePhysics, and Swaminathan Sundararaman, ParallelM
8:15 am–8:45 am
8:45 am–10:00 am
Learning @ the Edge
Session Chairs: Shadi A. Noghabi, UIUC/Microsoft Research, and Irfan Ahmad, CachePhysics
Xingzhou Zhang and Yifan Wang, Wayne State University, Institute of Computing Technology, CAS, University of Chinese Academy of Sciences; Weisong Shi, Wayne State University
Machine learning has changed the computing paradigm. Products today are built with machine intelligence as a central attribute, and consumers are beginning to expect near-human interaction with the appliances they use. However, much of the deep learning revolution has been limited to the cloud. Recently, several machine learning packages based on edge devices have been announced which aim to offload the computing to the edges. However, little research has been done to evaluate these packages on the edges, making it difficult for end users to select an appropriate pair of software and hardware. In this paper, we make a performance comparison of several state-of-the-art machine learning packages on the edges, including TensorFlow, Caffe2, MXNet, PyTorch, and TensorFlow Lite. We focus on evaluating the latency, memory footprint, and energy of these tools with two popular types of neural networks on different edge devices. This evaluation not only provides a reference to select appropriate combinations of hardware and software packages for end users but also points out possible future directions to optimize packages for developers.
Zeyi Tao and Qun Li, College of William and Mary
Training machine learning model on IoT device is a natural trend due to the growing computation power and the great ability to collect various data of modern IoT de- vice.In this work, we consider an edge based distributed deep learning framework in which many edge devices collaborate to train a model while using an edge server as the parameter server. However, the high network communication cost of synchronizing gradients and parameters between edge devices and cloud is a bottleneck. We propose a new method called edge Stochastic Gradient Descent (eSGD) for scaling up edge training of convolutional neural networks. eSGD is a family of sparse schemes with both convergence and practical performance guarantees. eSGD includes two mechanisms to improve the first order gradient based optimization of stochastic objective functions in edge scenario. First, eSGD determines which gradient coordinates are important and only transmits important gradient coordinates to cloud for synchronizing. This important update can aggressively reduce the communication cost. Second, momentum residual accumulation is designed for tracking out-of-date residual gradient coordinates to avoid low convergence rate caused by sparse updates. Our experiments show that we reach 91.2%, 86.7%, 81.5% accuracy on MNIST data set with gradient drop ratio 50%, 75%, 87.5% respectively.
Samuel S. Ogden and Tian Guo, Worcester Polytechnic Institute
In this paper, we propose a novel mobile deep inference platform, MODI, that delivers good inference performance. MODI improves deep learning powered mobile applications performance with optimizations in three complementary aspects. First, MODI provides a number of models and dynamically selects the best one during runtime. Second, MODI extends the set of models each mobile application can use by storing high quality models at the edge servers. Third, MODI manages a centralized model repository and periodically updates models at edge locations, ensuring up-to-date models for mobile applications without incurring high network latency. Our evaluation demonstrates the feasibility of trading off inference accuracy for improved inference speed, as well as the acceptable performance of edge-based inference.
Arun Ravindran and Anjus George, UNC Charlotte
Multi-camera real-time vision at the Edge is facilitated by low-latency distributed data stores. In this paper, we take the position that latency criticality in the challenging operating conditions at the Edge can only be attained through application specific designs incorporating autonomous computing techniques. In our initial prototype, we implement a key-value Edge data store that autonomously monitors run-time conditions to maintain latency-criticality of one class of data (feature vectors), while sacrificing the latency and accuracy of another class of data (keyframes). Early results show a median latency improvement of 84.8% over non-autonomous operation, for videos with large scene dynamics, and operational conditions of intermittent wireless channel interference.
10:00 am–10:15 am
Break with Refreshments
Essex Ballroom Foyer
10:15 am–10:45 am
10:45 am–12:00 pm
Security, Consistency, and Privacy
Session Chairs: Raj Nair, Ericsson, and Swaminathan Sundararaman, ParallelM
Ketan Bhardwaj, Joaquin Chung Miranda, and Ada Gavrilovska, Georgia Institute of Technology
Application-level DDoS attacks mounted using compromised IoT devices are emerging as a critical problem. The application-level and seemingly legitimate nature of traffic in such attacks renders most existing solutions ineffective, and the sheer amount and distribution of the generated traffic make mitigation extremely costly. This paper proposes a new approach which leverages edge computing infrastructure to accelerate the detection and the arrest of such attacks, limiting their damaging impact. Our preliminary investigation shows promise for up to 10x faster detection that reduces up to 82% of the Internet traffic due to IoT-DDoS.
Yunlong Mao, Nanjing University; Shanhe Yi and Qun Li, College of William & Mary; Jinghao Feng, Fengyuan Xu, and Sheng Zhong, Nanjing University
Deep convolutional neural networks (DNNs) have brought significant performance improvements to face recognition. However the training can hardly be carried out on mobile devices because the training of these models requires much computational power. An individual user with the demand of deriving DNN models from her own datasets usually has to outsource the training procedure onto a cloud or edge server. However this outsourcing method violates privacy because it exposes the users' data to curious service providers. In this paper, we utilize the differentially private mechanism to enable the privacy-preserving edge based training of DNN face recognition models. During the training, DNN is split between the user device and the edge server in a way that both private data and model parameters are protected, with only a small cost of local computations. We show that our mechanism is capable of training models in different scenarios, e.g., from scratch, or through fine-tuning over existed models.
Liangkai Liu and Xingzhou Zhang, Wayne State University; Mu Qiao, IBM Research-Almaden; Weisong Shi, Wayne State University
Ridesharing services, such as Uber and Didi, have enjoyed great popularity in our daily life. However, it remains a big challenge to guarantee the passenger and driver safety during the rides. In this paper, we propose an edge-based attack detection in ridesharing services, namely SafeShareRide, which can detect dangerous events happening on the vehicle in near real time. The detection of SafeShareRide consists of three stages: speech recognition, driving behavior detection and video capture and analysis. In our preliminary work, we implement the three detection stages by leveraging opensource algorithms and demonstrate the applicability of SafeShareRide. Furthermore, we identify several observations for smart phone based edge computing systems.
Seyed Hossein Mortazavi, University of Toronto; Bharath Balasubramanian, AT&T Labs-Research; Eyal de Lara, University of Toronto; Shankaranarayanan Puzhavakath Narayanan, AT&T Labs-Research
We describe a distributed datastore tailored for edge computing that provides session consistency between otherwise eventual consistent replicas. Existing solutions for session consistency assume that the client is associated with the same replica through a session. However, in edge computing use-cases, a client interacts with multiple replicas housed on different datacenters over time, either as a result of application partitioning, or client mobility. Our core algorithmic innovation is our client reconciliation algorithm that enforces session consistency by tracking and migrating only the client-affected keys between the replicas. Our results show that our approach provides session consistency at a fraction of the latency and bandwidth costs of a strongly consistent system, and with reasonable migration costs.
Zijiang Hao, Shanhe Yi, and Qun Li, College of William and Mary
Fast event ordering is critical for delay-sensitive edge computing applications that serve massive geographically distributed clients. Using a centralized cloud to determine the event order suffers from unsatisfactory latency. Naive edge-centric solutions, which designate one edge node to order all the events, have scalability and single point of failure issues. To address these problems, we propose EdgeCons, a novel consensus algorithm optimized for edge computing networks. EdgeCons achieves fast consensus by running a sequence of Paxos instances among the edge nodes and dynamically distributing their leadership based on the recent running history. It also guarantees progressiveness by incorporating a reliable, backend cloud. A preliminary evaluation shows that EdgeCons works more efficiently than the state-of-the-art consensus algorithms, in the context of achieving fast event ordering in edge computing networks.
12:00 pm–1:30 pm
Luncheon for Workshop Attendees
Essex Ballroom South
2:00 pm–3:15 pm
Infrastructure for Edge Computing
Session Chairs: Ketan Bharadwaj, Georgia Institute of Technology, and Bharath Balasubramanian, ATT Labs Research
Chi-Yu Li, Hsueh-Yang Liu, Po-Hao Huang, and Hsu-Tung Chien, National Chiao Tung University; Guan-Hua Tu, Michigan State University; Pei-Yuan Hong and Ying-Dar Lin, National Chiao Tung University
Low-latency demands for cellular networks have attracted much attention. Mobile edge computing (MEC), which deploys a cloud computing platform at the edge closer to mobile users, has been introduced as an enabler of low-latency performance in 4G and 5G networks. In this paper, we propose an MEC platform deployment solution in 4G LTE networks using a middlebox approach. It is standard-compliant and transparent to existing cellular network components, so they need not be modified. The MEC middlebox sits on the S1 interface, which connects an LTE base station to its core network, and does traffic filtering, manipulation and forwarding. It enables the MEC service for mobile users by hosting application servers. Such middlebox approach can save deployment cost and be easy to install. It is different from other studies that require modifications on base stations or/and core networks. We have confirmed its viability through a prototype based on the OpenAirInterface cellular platform.
Christopher Meiklejohn, Universite catholique de Louvain; Heather Miller, Northeastern University; Zeeshan Lakhani, Comcast Cable
Edge computing promises lower latency interactions for clients operating at the edge by shifting computation away from Data Centers to Points of Presence which are more abundant and located geographically closer to end users. However, most commercially available infrastructure for edge computing focuses on applications without shared state. In this paper, we present the Red Wedding Problem, a real-world scenario motivating the need for stateful computations at the edge. We sketch the design and implementation of a prototype database for operation at the edge that addresses the issues presented in the Red Wedding Problem and present issues around implementing our solution on commercial edge infrastructure due to limitations in these offerings.
Youdong Chen and Qiangguo Feng, School of Mechanical Engineering and Automation, Beihang University, Beijing 100191, China; Weisong Shi, Department of Computer Science, Wayne State University, USA
As more sensors and actuators are deployed in industrial manufacturing, the industry requires a new production system architecture that offers better real-time and network transmission performance. Yet cloud computing (based on a centralized datacenter) is limited in its possibilities, because it suffers from heavy bandwidth costs and lengthy time delays. As a solution, we propose an industrial robot system based on edge computing. Here, we present its three-layer architecture in detail: the cloud, edge, and physical resource layers. Initially, we deploy an edge node near the data sources, to integrate various devices’ interfaces and acts as a raw data filter. Then, we apply the proposed system on the robotic welding of the membrane wall cell. Finally, we test the system by conducting an experiment. The results demonstrate the system’s feasibility and prove that the system yields better real-time and network transmission performance than a cloud-based scenario.
Edge Computing Resource Management System: a Critical Building Block! Initiating the debate via OpenStack
Ronan-Alexandre Cherrueau, Adrien Lebre, and Dimitri Pertin, STACK Research Group - IMT-Atlantique, Inria, LS2N, France; Fetahi Wuhib, Ericsson Research, Canada; João Monteiro Soares, Ericsson Research, Sweden
While it is clear that edge infrastructures are required for emerging use-cases related to IoT, VR or NFV, there is currently no resource management system able to deliver all features for the edge that made cloud computing successful (e.g., an OpenStack for the edge). Since building a system from scratch is seen by many as impractical, this paper provides reflections regarding how existing solutions can be leveraged. To that end, we provide a list of the features required to operate and use edge computing resources, and investigate how an existing IaaS manager (i.e., OpenStack) satisfies these requirements. Finally, we identify from this study two approaches to design an edge infrastructure manager that fulfils our requirements, and discuss their pros and cons. This paper aims at initiating the discussion in our community.
Madhumita Bharde, Annmary Justine K, Suparna Bhattacharya, and Dileep Deepa Shree, Hewlett Packard Enterprise
Powerful edge compute frameworks address the issue of latency for IoT data processing at the edge. However, continuous application layer WAN streaming to core for consolidated deep analysis and learning consumes excessive bandwidth and becomes the bottleneck for responsiveness at the core. A state-of-the-art storage or hyperconverged system, on the other hand, advertises compelling in-built features like WAN efficient data protection and delta replication, global unified management, space and bandwidth saving through inline data compression and deduplication. Traditional storage semantics and services, however, are built for data at rest while edge analytics prioritizes responsiveness by processing and moving streaming data at an application layer. In this paper, we propose to enable streaming of IoT data transparently through storage replication. Based on this foundation, we further present light-weight storage plugins to reduce IoT data transfer by detecting and translating semantic redundancies to a deduplication friendly form. Our early results demonstrate that (a) leveraging storage to take responsibility of streaming data in an application consistent way results in efficient data transfer (b) real-world IoT time-series datasets exhibit a high degree of similarity which can be detected to reduce data transfer from edge (c) video streams for autonomous cars, the transfer of which cannot be reduced enough using traditional video compression or storage deduplication techniques, have significant semantic redundancy. Collectively, advancing research in this direction paves the way to enhance the versatility of state-of-the-art infrastructure for optimized edge computing.
3:15 pm–3:30 pm
Break with Refreshments
Essex Ballroom Foyer
3:30 pm–5:00 pm
Session Chairs: Padmanabhan Pillai, Intel, and Jussi Kangasharju, University of Helsinki
Shizhe Zang, Wei Bao, Phee Lep Yeoh, Branka Vucetic, and Yonghui Li, University of Sydney
Due to the strong potential of edge computing, service providers are aware of its excellent market opportunity. In this paper, we examine new pricing plans for edge-computing services that jointly consider communication and computing costs in mobile devices and edge servers. In our proposed model, users can freely choose to use: 1) a pay-as-you-go payment for communication and computing costs, 2) an upfront data plan for unlimited communication volume, 3) an upfront computing plan for unlimited computing workload, or 4) a combo plan for unlimited communication volume and computing workload. Based on this, we design a smart plan-purchasing scheme, namely smart online reservation (SOR) scheme for users and prove it achieves the best possible competitive ratio among all possible online deterministic schemes. With SOR adopted by users, we develop pricing plan guidelines for service providers to achieve profitability by motivating users to purchase the combo plan. We provide trace-driven simulations to verify the advantage of SOR and identify how different prices influence user behavior.
Tim Süß, Tunahan Kaya, Markus Mäsker, and André Brinkmann, Johannes Gutenberg University Mainz
The availability and usage of embedded systems increases permanently and the industry drives the IoT to become more and more relevant in daily life. Factory lines, planes and cars, traffic lights, or even clothes are equipped with sensors and small computers constantly communicating with the outside world. One challenge in maintaining those devices is updating their software. Due to slow connections or only because of the huge amount of devices data transfers can be problematic. Data compression algorithms can be applied to reduce the amount of data that must be transferred. A data reduction technique that provides high efficiency, but which has not been considered so far for embedded systems is data deduplication. In this work we present the results of a long term study for updating a car multimedia system. The results show that deduplication can achieve significantly better results than commonly used data compression techniques.
Srikumar Venugopal, Michele Gazzetti, Yiannis Gkoufas, and Kostas Katrinis, IBM Research
Extracting value from insights on unstructured data on the Internet of Things and Humans is a major trend in capitalizing on digitization. To date, the design space for doing AI inference on the edge has been highly binary: either consuming cloud-based inference services through edge APIs or running full-fledged deep models on edge devices. In this paper, we break this design space duality by proposing the Semantic Cache, an approach that blends best-of-breed features of the extreme ends of the current design space. Early evaluation results on a first prototype implementation of our semantic cache service on object classification tasks shows tremendous inference latency reduction, when compared to cloud-only inference, and high potential in scoring adequate accuracy for a plurality of AI use-cases.
Ioannis Psaras, Onur Ascigil, Sergi Rene, and George Pavlou, UCL, UK; Alex Afanasyev, FIU, USA; Lixia Zhang, UCLA, USA
In a future IoT-dominated environment the majority of data will be produced at the edge, which may be moved to the network core. We argue that this reverses today’s “core-to-edge” data flow to an “edge-to-core” model and puts severe stress on edge access/cellular links. In this paper, we propose a data-centric communication approach which treats storage and wire the same as far as their ability to supply the requested data is concerned. Given that storage is cheaper and scales easier than wires, we argue for enhancing network connectivity with local storage services (e.g., in WiFi Access Points, or similar) at the edge of the network. Such local storage services can be used to temporarily store IoT and user-generated data at the edge, prior to data-cloud synchronization.
Harshit Gupta, Zhuangdi Xu, and Umakishore Ramachandran, Georgia Institute of Technology
The increasing adoption of the Internet of Things (IoT) paradigm in structuring services for “smart cities” has led to an explosion of data generated, ingested, processed, and stored in the system. Traditional time-series databases being Cloud-based are not suitable for meeting the low-latency requirement of such geo-distributed smart services. We propose DataFog, a geo-distributed data-management platform at the edge of the network to cater to the needs of smart services for the IoT age. We identify key challenges towards building such a database over a widely geo-distributed and heterogeneous edge computing environment. Preliminary evaluations show the performance potential of DataFog in comparison to state-of-the-art distributed data-stores. DataFog is the first system, to the best of the authors’ knowledge, that is meant for data-management at the network edge.
Saman Biookaghazadeh, Ming Zhao, and Fengbo Ren, Arizona State University
The rapid growth of Internet-of-things (IoT) and artificial intelligence applications have called forth a new computing paradigm--edge computing. In this paper, we study the suitability of deploying FPGAs for edge computing from the perspectives of throughput sensitivity to workload size, architectural adaptiveness to algorithm characteristics, and energy efficiency. This goal is accomplished by conducting comparison experiments on an Intel Arria 10 GX1150 FPGA and an Nvidia Tesla K40m GPU. The experiment results suggest that the key advantages of adopting FPGAs for edge computing over GPUs are three-fold: 1) FPGAs can provide a consistent throughput invariant to the size of application workload, which is critical to aggregating individual service requests from various IoT sensors; (2) FPGAs offer both spatial and temporal parallelism at a fine granularity and a massive scale, which guarantees a consistently high performance for accelerating both high-concurrency and high-dependency algorithms; and (3) FPGAs feature 3--4 times lower power consumption and up to 30.7 times better energy efficiency, offering better thermal stability and lower energy cost per functionality.
5:00 pm–6:00 pm
Poster Session and Happy Hour
Sponsored by Ericsson
Join your colleagues over complimentary food and drinks and check out the cool new ideas and the latest preliminary research on display at the Poster Session and Happy Hour.