International audience

Edge computing places the computational services and resources closer to the user proximity, to reduce latency, and ensure the quality of service and experience. Low latency, context awareness and mobility support are the major contributors to edge-enabled smart systems. Such systems require handling new situations and change on the fly and ensuring the quality of service while only having access to constrained computation and communication resources and operating in mobile, dynamic and ever-changing environments. Hence, adaptation and self-organisation are crucial for such systems to maintain their performance, and operability while accommodating new changes in their environment. This article reviews the current literature in the field of adaptive edge computing systems. We use a widely accepted taxonomy, which describes the important aspects of adaptive behaviour implementation in computing systems. This taxonomy discusses aspects such as adaptation reasons, the various levels an adaptation strategy can be implemented, the time of reaction to a change, categories of adaptation technique and control of the adaptive behaviour. In this article, we discuss how these aspects are addressed in the literature and identify the open research challenges and future direction in adaptive edge computing systems. The results of our analysis show that most of the identified approaches target adaptation at the application level, and only a few focus on middleware, communication infrastructure and context. Adaptations that are required to address the changes in the context, changes caused by users or in the system itself are also less explored. Furthermore, most of the literature has opted for reactive adaptation, although proactive adaptation is essential to maintain the edge computing systems’ performance and interoperability by anticipating the required adaptations on the fly. Additionally, most approaches apply a centralised adaptation control, which does not perfectly fit the mostly decentralised/distributed edge computing settings.

End users can mutually share their computing resources in ad-hoc computing environments with code offloading. This augments the computational power of resource-constrained mobile devices and enables interactive user-facing applications that would otherwise exceed single device capabilities. However, ad-hoc computing comes along with new challenges such as heterogeneity and unreliability of devices. Resource consumers have to make task scheduling decisions without relying on a centralized scheduler to facilitate sub-second response times in environments with communication latencies that are in the order of the task execution times. In this paper, we present a decentralized low-latency task scheduling approach that minimizes job execution times in heterogeneous ad-hoc environments. We propose two decentralized task scheduling algorithms that select powerful computing resources for parallel task execution while avoiding delays that arise from congested devices. We provide an analytical model of the performance of these algorithms before conducting an extensive evaluation based on real-world applications and a realistic computing infrastructure. Our results show that decentralized scheduling can dynamically adapt to varying system load and outperform a central scheduler in both task and job execution times, which enables low-latency task offloading in ad-hoc environments.

Computationally intensive tasks of IoT applications can be offloaded to powerful devices in the edge. Code offloading reduces energy consumption and increases performance. However, applications that use face recognition, machine learning, or image rendering, rely on large amounts of data. The transfer of this data leads to latencies which contradicts the responsiveness required by many pervasive applications. As a solution, decoupling the data from the tasks allows to apply new scheduling strategies that place data on remote devices before the actual task execution. Grid computing approaches use this technique effectively, however, edge computing introduces further challenges such as device fluctuation and heterogeneity.In this paper, we propose a data management approach for edge computing environments that decouples data placement from task scheduling. We present a multi-level scheduler, which places data on resource providers in the system considering multiple context dimensions. The scheduler allocates tasks according to the current context and observes the state during runtime. If required, the system adjusts the number of data copies to optimize the trade-off between execution latencies and data management overhead. The paper has three contributions: (1) a context-aware multi-level scheduler, (2) the integration of four data placement, three task scheduling, and three runtime adaptation algorithms, (3) an evaluation in a real-world testbed.

In the past decade, cloud computing has shown its potential to provide powerful and reliable resources at the core of the network. Many applications can benefit from the wide range of cloud services. However, as applications in the Internet of Things become more common, the computing environment faces new requirements and challenges that cloud computing cannot meet. Fog and edge computing paradigms can fill this gap by moving computation from the core to the edge of the network. While multiple solutions for edge- centric networks have been proposed, there is still confusion about the terminology and classification of edge-centric architectures. In this paper, we summarize the current discussion about fog and edge computing systems. Further, we identify application areas of these systems in the Internet of Things.

In edge computing systems, computation is rather offloaded to nearby resources than to the cloud, due to latency reasons. However, the performance demand in the edge grows steadily, which makes nearby resources insufficient for many applications. Additionally, the amount of parallel tasks in the edge increases, based on trends like machine learning, Internet of Things, and artificial intelligence. This introduces a trade- off between the performance of the cloud and the communication latency of the edge. However, many edge devices have powerful co-processors in form of their graphics-processing unit (GPU), which are mostly unused. These processing units have specialized parallel architectures, which are different from standard CPUs and complex to use. In this paper, we present GPU-accelerated task execution for edge computing environments. The paper has four contributions. First, we design and implement a GPU system extension for our Tasklet system - a distributed computing system, which supports edge- and cloud-based task offloading. Second, we introduce a computational abstraction for GPUs in form of a virtual machine, which exploits parallelism while considering device heterogeneity and maintaining unobtrusiveness. Third, we offer an easy-to-use programming interface for the rather complex architecture of GPUs. Fourth, we evaluate our prototype in a real- world testbed and compare the GPU performance to standard edge resources.

Today's modern computing landscape consists of a huge amount of heterogeneous devices, including powerful, stable desktop computers as well as lightweight, unreliable mobile edge devices. This heterogeneity in terms of computation power and reliability increases the complexity for fault tolerance in distributed computing systems. When tasks are offloaded, slow resource providers easily become the bottleneck of a parallel computation. Further, unstable edge devices can leave the system spontaneously, discontinue remote tasks executions, and therefore lose the computation progress. These two effects increase the response time for remote task executions. In this paper, we introduce two mechanisms to avoid delayed or lost task executions caused by edge devices. This paper has five contributions. First, we define a failure model and identify the parameters that determine the magnitude of delays caused by faults and performance bottlenecks. Second, we present reactive and proactive task migration to handle system leaves. Third, we show how computational bottlenecks can be avoided by two-dimensional context-aware task partitioning. Fourth, we integrate these two solutions into an existing heterogeneous distributed computing system. Fifth, we run an evaluation on a real- world testbed to show the benefits of the solutions in practice. The evaluation shows, that we can improve systems with device fluctuation and heterogeneity by up to 39% and 53% respectively.

Mobile devices are ubiquitous but their resources are limited. However, they must be capable to run computationally intensive software, for example for image stitching, face recognition, and simulation-based artificial intelligence. As a solution, mobile devices can use nearby resources to offload computation. Distributed computing environments provide such features but ignore the nature of mobile devices, such as mobility, network, or battery changes. This leads to long delays, which reduce the quality of experience for the user. In this paper, we present Mobile Tasklets, a mobile extension of our distributed computing middleware. The design of Mobile Tasklets includes context monitoring, context-aware scheduling mechanisms, and an Android API for application integration. We identify the challenges of the integration of mobile devices into our distributed computing environment. We evaluate Mobile Tasklets in a realworld testbed with different context settings.

Scheduling in distributed computing systems is the process of allocating resources to a computational task. The complexity of this allocation process increases with the amount of criteria that are considered for the scheduling decision. Pervasive computing systems show a high degree of heterogeneity and dynamism. The constant joining and leaving of devices makes the system error-prone and less predictable. The involved devices differ in various properties that we subsume as their context. We argue, that these context dimensions can be used to implement fault-avoidant scheduling strategies. In this paper, we introduce the concept of context-aware scheduling for pervasive computing systems. The schedulers in these systems consider multiple context dimensions to avoid failing resource providers. We discuss relevant context dimensions, develop context-aware scheduling strategies and implement them into an existing distributed computing system. We show how to monitor the context dimensions and evaluate the fault-avoidant scheduling strategies in a large-scale simulation.

The modern computing landscape consists of numerous heterogeneous devices, all of which can contribute to a distributed environment as generic computation resources. In unstructured environments, resources can easily be shared and consumed at the cost of certainty. While some applications can handle such a best-effort service, many others require execution qualities, e.g., reliability or speed. We introduce Quality of Computation (QoC) as a thin layer on top of uniformly abstracted best-effort resources, which allows to tailor computation tasks to application-specific needs. The QoC layer provides execution guarantees for reliability, speed, precision, privacy, cost, and energy. We demonstrate QoC on the basis of the Tasklet system. Tasklets are fine-grained units of computation that can be issued for local or remote execution. The Tasklet system has two layers. Below is the best-effort execution layer, a virtual machine that provides raw computation. Above is the orchestration layer, which federates these virtual machines to one distributed computing environment and enforces the mechanisms that guarantee the requested QoC. We evaluated the performance of QoC in the Tasklet system in various scenarios. Results indicate that our system provides QoC guarantees at minimal performance cost.

Pervasive applications are designed to support users in their daily lives. In order to provide their services, these applications interact with the environment, i.e. their context. They either adapt themselves as a reaction to context changes, or adapt the context via actuators according to their needs. If multiple applications are executed in the same context, interferences are likely to occur. In this paper, we present COMITY-a framework for interference management in multi-platform pervasive systems. Based on contracts specifying an application's interaction with the context, the framework automatically detects interferences and resolves them through a coordinated application adaptation. We analyze the problem of interference resolution, discuss respective algorithms and extensively evaluate our prototype.

This paper establishes weak and strong universal consistency of regression estimates based on normalized radial basis function networks when the network parameters are chosen by empirical risk minimization.

This paper presents relative stability properties of various nonparametric density estimators (histogram, kernel estimates) and of regression estimators (partitioning, kernel, and nearest neighbor estimates). In density estimation, let En denote the L/sub 1/ error of an estimate calculated from n data, whereas in regression estimation, the L/sub 2/ error of the estimate is used. Sufficient conditions for E/sub n//E{E/sub n/}/spl rarr/1 in probability are provided. If this limit holds, the asymptotic behavior of the random error E/sub n/ can be characterized by its expectation E{E/sub n/},, and one may apply, for example, the established rate-of-convergence results for E{En}.

An estimator E/spl circ/(d/sub n/,n) of the conditional expectation E[X/sub n+1/|X/sub n/,...,X(n-d/sub n/+1)] in a centered, stationary, and ergodic Gaussian process {X/sub i/}/sub i/ with absolutely summable Wold coefficients is constructed on the basis of having observed X/sub 1/,...,X/sub n/. For a suitable choice of the length dn/spl rarr//spl infin/ (n/spl rarr//spl infin/) of the past covered by the conditional expectation, it is established that |E/spl circ/(d/sub n/,n)-E[X/sub n+1/|X/sub n/,...,X(n-d/sub n/+1)]|/spl rarr/0 with probability 1. In addition, sufficient conditions for |E[X/sub n+1/|X/sub n/,X/sub n-1/,...]-E[X/sub n+1/|X/sub n/,...,X(n-d/sub n/+1)]| /spl rarr/0 to hold with probability 1 are given, that is, conditions under which E/spl circ/(d/sub n/,n) can be used as a strongly consistent forecaster for |E[X/sub n+1/|X/sub n/,X/sub n-1/,...].