Doctor of Philosophy / Computing and Information Sciences / Daniel A. Andresen / The need for increased performance of mobile device directly conflicts with the desire for longer battery life. Offloading computation to resourceful servers is an effective method to reduce energy consumption and enhance performance for mobile applications. Today, most mobile devices have fast wireless link such as 4G and Wi-Fi, making computation offloading a reasonable solution to extend battery life of mobile device. Android provides mechanisms for creating mobile applications but lacks a native scheduling system for determining where code should be executed. We present Jade, a system that adds sophisticated energy-aware computation offloading capabilities to Android applications. Jade monitors device and application status and automatically decides where code should be executed. Jade dynamically adjusts offloading strategy by adapting to workload variation, communication costs, and device status. Jade minimizes the burden on developers to build applications with computation offloading ability by providing easy-to-use Jade API. Evaluation shows that Jade can effectively reduce up to 37% of average power consumption for mobile device while improving application performance.

Edge computing is a distributed computing paradigm where computing resources are located physically closer to the data source compared to the traditional cloud computing paradigm. Edge computing enables computation offloading from resource-constrained devices to more powerful servers in the edge and cloud. To offer edge and cloud support to real-time industrial applications, the communication to the servers and the server-side computation needs to be predictable. However, the predictability of offloading cannot be guaranteed in an environment where multiple devices are competing for the same edge and cloud resources due to potential server-side scheduling conflicts. To the best or our knowledge, no offloading scheme has been proposed that provides a highly predictable real-time task scheduling in the face of multiple devices offloading to a set of heterogeneous edge/cloud servers. Hence, this thesis approaches the problem of predictable offloading in real-time environments by proposing a centralized server time reservation system to schedule the offloading of real-time tasks to edge and cloud servers. Our reservation system allows end-devices to request external execution time in advance for real-time tasks that will be generated in the future, therefore when such a task is created, it already has a designated offloading server that guarantees its timely execution. Furthermore, this centralized reservation system is capable of optimizing the reservation scheduling strategy with the goal of minimizing energy consumption of edge servers while meeting the stringent deadline constraints of real-time applications. / Edge computing är ett distribuerat datorparadigm där datorresurser är fysiskt placerade närmare datakällan jämfört med det traditionella molnberäkningsparadigmet. Edge computing möjliggör beräkningsavlastning från resursbegränsade enheter till mer kraftfulla servrar i kanten och molnet. För att erbjuda kant- och molnstöd till industriella tillämpningar i realtid måste kommunikationen till servrarna och beräkningen på serversidan vara förutsägbar. Förutsägbarheten av avlastning kan dock inte garanteras i en miljö där flera enheter konkurrerar om samma kant- och molnresurser på grund av potentiella schemaläggningskonflikter på serversidan. Så vitt vi vet har inget avlastningsschema föreslagits som ger en mycket förutsägbar uppgiftsschemaläggning i realtid inför flera enheter som laddas av till en uppsättning heterogena edge-/molnservrar. Därför närmar sig denna avhandling problemet med förutsägbar avlastning i realtidsmiljöer genom att föreslå ett centraliserat servertidsreservationssystem för att schemalägga avlastningen av realtidsuppgifter till edge- och molnservrar. Vårt reservationssystem tillåter slutenheter att begära extern exekveringstid i förväg för realtidsuppgifter som kommer att genereras i framtiden, därför när en sådan uppgift skapas har den redan en utsedd avlastningsserver som garanterar att den utförs i tid. Dessutom kan detta centraliserade bokningssystem optimera bokningsschemaläggningsstrategin med målet att minimera energiförbrukningen för edge-servrar samtidigt som de stränga deadline-begränsningarna för realtidsapplikationer uppfylls.

Le délestage de calcul ou de code est une technique qui permet à un appareil mobile avec une contrainte de ressources d'exécuter à distance, entièrement ou partiellement, une application intensive en calcul dans un environnement Cloud avec des ressources suffisantes. Le délestage de code est effectué principalement pour économiser de l'énergie, améliorer les performances, ou en raison de l'incapacité des appareils mobiles à traiter des calculs intensifs. Plusieurs approches et systèmes ont été proposés pour délester du code dans le Cloud tels que CloneCloud, MAUI et Cyber Foraging. La plupart de ces systèmes offrent une solution complète qui traite différents objectifs. Bien que ces systèmes présentent en général de bonnes performances, un problème commun entre eux est qu'ils ne sont pas adaptés aux applications temps réel telles que les jeux vidéo, la réalité augmentée et la réalité virtuelle, qui nécessitent un traitement particulier. Le délestage de code a connu un récent engouement avec l'avènement du MEC et son évolution vers le edge à multiple accès qui élargit son applicabilité à des réseaux hétérogènes comprenant le WiFi et les technologies d'accès fixe. Combiné avec l'accès mobile 5G, une pléthore de nouveaux services mobiles apparaîtront, notamment des service type URLLC et eV2X. De tels types de services nécessitent une faible latence pour accéder aux données et des capacités de ressources suffisantes pour les exécuter. Pour mieux trouver sa position dans une architecture 5G et entre les services 5G proposés, le délestage de code doit surmonter plusieurs défis; la latence réseau élevée, hétérogénéité des ressources, interopérabilité des applications et leur portabilité, la consommation d'énergie, la sécurité, et la mobilité, pour citer quelques uns. Dans cette thèse, nous étudions le paradigme du délestage de code pour des applications a temps réel, par exemple; les jeux vidéo sur équipements mobiles et le traitement d'images. L'accent sera mis sur la latence réseau, la consommation de ressources, et les performances accomplies. Les contributions de la thèse sont organisées sous les axes suivants : Étudier le comportement des moteurs de jeu sur différentes plateformes en termes de consommation de ressources (CPU / GPU) par image et par module de jeu ; Étudier la possibilité de distribuer les modules du moteur de jeu en fonction de la consommation de ressources, de la latence réseau, et de la dépendance du code ; Proposer une stratégie de déploiement pour les fournisseurs de jeux dans le Cloud, afin de mieux exploiter les ressources, en fonction de la demande variable en ressource par des moteurs de jeu et de la QoE du joueur ; Proposer une solution de délestage statique de code pour les moteurs de jeu en divisant la scène 3D en différents objets du jeu. Certains de ces objets sont distribués en fonction de la consommation de ressources, de la latence réseau et de la dépendance du code ; Proposer une solution de délestage dynamique de code pour les moteurs de jeu basée sur une heuristique qui calcule pour chaque objet du jeu, le gain du délestage. En fonction de ce gain, un objet peut être distribué ou non ; Proposer une nouvelle approche pour le délestage de code vers le MEC en déployant une application sur la bordure du réseau (edge) responsable de la décision de délestage au niveau du terminal et proposer deux algorithmes pour prendre la meilleure décision concernant les tâches à distribuer entre le terminal et le serveur hébergé dans le MEC. / Computation offloading is a technique that allows resource-constrained mobile devices to fully or partially offload a computation-intensive application to a resourceful Cloud environment. Computation offloading is performed mostly to save energy, improve performance, or due to the inability of mobile devices to process a computation heavy task. There have been a numerous approaches and systems on offloading tasks in the classical Mobile Cloud Computing (MCC) environments such as, CloneCloud, MAUI, and Cyber Foraging. Most of these systems are offering a complete solution that deal with different objectives. Although these systems present in general good performance, one common issue between them is that they are not adapted to real-time applications such as mobile gaming, augmented reality, and virtual reality, which need a particular treatment. Computation offloading is widely promoted especially with the advent of Mobile Edge Computing (MEC) and its evolution toward Multi-access Edge Computing which broaden its applicability to heterogeneous networks including WiFi and fixed access technologies. Combined with 5G mobile access, a plethora of novel mobile services will appear that include Ultra-Reliable Low-latency Communications (URLLC) and enhanced Vehicle-toeverything (eV2X). Such type of services requires low latency to access data and high resource capabilities to compute their behaviour. To better find its position inside a 5G architecture and between the offered 5G services, computation offloading needs to overcome several challenges; the high network latency, resources heterogeneity, applications interoperability and portability, offloading frameworks overhead, power consumption, security, and mobility, to name a few. In this thesis, we study the computation offloading paradigm for real-time applications including mobile gaming and image processing. The focus will be on the network latency, resource consumption, and accomplished performance. The contributions of the thesis are organized on the following axes : Study game engines behaviour on different platforms regarding resource consumption (CPU/GPU) per frame and per game module; study the possibility to offload game engine modules based on resource consumption, network latency, and code dependency ; propose a deployment strategy for Cloud gaming providers to better exploit their resources based on the variability of the resource demand of game engines and the QoE ; propose a static computation offloading-based solution for game engines by splitting 3D world scene into different game objects. Some of these objects are offloaded based on resource consumption, network latency, and code dependency ; propose a dynamic offloading solution for game engines based on an heuristic that compute for each game object, the offloading gain. Based on that gain, an object may be offloaded or not ; propose a novel approach to offload computation to MEC by deploying a mobile edge application that is responsible for driving the UE decision for offloading, as well as propose two algorithms to make best decision regarding offloading tasks on UE to a server hosted on the MEC.

The emergence of 5G alongside the distributed computing paradigm called Edge computing has prompted a tremendous change in the industry through the opportunity for reducing network latency and energy consumption and providing scalability. Edge computing extends the capabilities of users’ resource-constrained devices by placing data centers at the edge of the network. Computation offloading enables edge computing by allowing the migration of users’ tasks to edge servers. Deciding whether it is beneficial for a mobile device to offload a task and on which server to offload, while environmental variables, such as availability, load, network quality, etc., are changing dynamically, is a challenging problem that requires careful consideration to achieve better performance. This project focuses on proposing lightweight and efficient algorithms to take offloading decisions from the mobile device perspective to benefit the user. Subsequently, heuristic techniques have been examined as a way to find quick but sub-optimal solutions. These techniques have been combined with a Multi-Armed Bandit algorithm, called Discounted Upper Confidence Bound (DUCB) to take optimal decisions quickly. The findings indicate that these heuristic approaches cannot handle the dynamicity of the problem and the DUCB provides the ability to adapt to changing circumstances without having to keep adding extra parameters. Overall, the DUCB algorithm performs better in terms of local energy consumption and can improve service time most of the times. / Utvecklingen av 5G har skett parallellt med det distribuerade beräkningsparadigm som går under namnet Edge Computing. Lokala datacenter placerade på kanten av nätverket kan reducera nätverkslatensen och energiförbrukningen för applikationer. Exempelvis kan användarenheter med begränsade resurser ges utökande möjligheter genom avlastning av beräkningsintensiva uppgifter. Avlastningen sker genom att migrera de beräkningsintensiva uppgifterna till en dator i datacentret på kanten. Det är dock inte säkert att det alltid lönar sig att avlasta en beräkningsintensiv uppgift från en enhet till kanten. Detta måste avgöras från fall till fall. Att avgöra om och när det lönar sig är ett svårt problem då förutsättningar som tillgänglighet, last, nätverkskvalitét, etcetera hela tiden varierar. Fokus i detta projekt är att identifiera enkla och effektiva algoritmer som kan avgöra om det lönar sig för en användare att avlasta en beräkningsintensiv uppgift från en mobil enhet till kanten. Heuristiska tekniker har utvärderats som en möjlig väg att snabbt hitta lösningar även om de råkar vara suboptimala. Dessa tekniker har kombinerats med en flerarmad banditalgoritm (Multi-Armed Bandit), kallad Discounted Upper Confidence Bound (DUCB), för att ta optimala beslut snabbt. Resultaten indikerar att dessa heuristiska tekniker inte kan hantera de dynamiska förändringar som hela tiden sker samtidigt som DUCB kan anpassa sig till dessa förändrade omständigheter utan att man måste addera extra parametrar. Sammantaget, ger DUCM-algoritmen bättre resultat när det gäller lokal energikonsumtion och kan i de flesta fallen förbättra tiden för tjänsten.

Um Systeme der fünften Generation zellularer Kommunikationsnetze (5G) zu ermöglichen, sind Energie effiziente Architekturen erforderlich, die eine zuverlässige Serviceplattform für die Bereitstellung von 5G-Diensten und darüber hinaus bieten können. Device Enhanced Edge Computing ist eine Ableitung des Multi-Access Edge Computing (MEC), das Rechen- und Speicherressourcen direkt auf den Endgeräten bereitstellt. Die Bedeutung dieses Konzepts wird durch die steigenden Anforderungen von rechenintensiven Anwendungen mit extrem niedriger Latenzzeit belegt, die den MEC-Server allein und den drahtlosen Kanal überfordern. Diese Dissertation stellt ein Berechnungs-Auslagerungsframework mit Berücksichtigung von Energie, Mobilität und Anreizen in einem gerätegestützten MEC-System mit mehreren Benutzern und mehreren Aufgaben vor, das die gegenseitige Abhängigkeit der Aufgaben sowie die Latenzanforderungen der Anwendungen berücksichtigt. / To enable fifth generation cellular communication network (5G) systems, energy efficient architectures are required that can provide a reliable service platform for the delivery of 5G services and beyond. Device Enhanced Edge Computing is a derivative of Multi-Access Edge Computing (MEC), which provides computing and storage resources directly on the end devices. The importance of this concept is evidenced by the increasing demands of ultra-low latency computationally intensive applications that overwhelm the MEC server alone and the wireless channel. This dissertation presents a computational offloading framework considering energy, mobility and incentives in a multi-user, multi-task device-based MEC system that takes into account task interdependence and application latency requirements.

State estimation filters are computationally intensive blocks used to calculate uncertain/unknown state values from noisy/not available sensor inputs in any autonomous systems. The inputs to the actuators depend on these filter’s output and thus the scheduling of filter has to be at very small time intervals. The aim of this thesis is to investigate the possibility of using hardware accelerators to perform this computation. To make a comparative study, 3 filters that predicts 4, 8 and 16 state information was developed and implemented in Arm real time and application purpose CPU, NVIDIA Quadro and Turing GPU, and Xilinx FPGA programmable logic. The execution, memory transfer time, and the total developement time to realise the logic in CPU, GPU and FPGA is discussed. The CUDA developement environment was used for the GPU implementation and Vivado HLS with SDSoc environment was used for the FPGA implementation. The thesis concludes that a hardware accelerator is needed if the filter estimates 16 or more state information even if the processor is entirely dedicated for the computation of filter logic. Otherwise, for a 4 and 8 state filter the processor shows similar performance as an accelerator. However, in a real time environment the processor is the brain of the system, so it has to give instructions to many other functions parallelly. In such an environment, the instruction and data caches of the processor will be disturbed and there will be a fluctuation in the execution time of the filter for every iteration. For this, the best and worst case processor timings are calculated and discussed. / Tillståndsberäkningsfilter är beräkningsintensiva block som används för att beräkna osäkra / okända tillståndsvärden från bullriga / ej tillgängliga sensoringångar i autonoma system. Ingångarna till manöverdonen beror på filterens utgång och därför måste schemaläggningen av filtret ske med mycket små tidsintervall. Syftet med denna avhandling är att undersöka möjligheten att använda hårdvaruacceleratorer för att utföra denna beräkning. För att göra en jämförande studie utvecklades och implementerades 3 filter som förutsäger information om 4, 8 och 16 tillstånd i realtid med applikationsändamålen CPU, NVIDIA Quadro och Turing GPU, och Xilinx FPGA programmerbar logik. Exekvering, minnesöverföringstid och den totala utvecklingstiden för att förverkliga logiken i båda hårdvarorna diskuteras. CUDAs utvecklingsmiljö användes för GPU-implementeringen och Vivado HLS med SDSoc-miljö användes för FPGA-implementering. Avhandlingen drar slutsatsen att en hårdvaru-accelerator behövs om filtret uppskattar information om mer än 16 tillstånd även om processorn är helt dedikerad för beräkning av filterlogik. För 4 och 8 tillståndsfilter, visar processorn liknande prestanda som en accelerator. Men i realtid är processorn hjärnan i systemet; så den måste ge instruktioner till många andra funktioner parallellt. I en sådan miljö kommer processorns instruktioner och datacacher att störas och det kommer att bli en fluktuation i exekveringstiden för filtret för varje iteration. För detta beräknas och diskuteras de bästa och värsta fallstiderna.

L'hétérogénéité et l'énergie sont deux considérations fondamentales pour les environnements informatiques ubiquitaires. Dans cette thèse, nous présentons notre approche pour gérer l'hétérogénéité et pour économiser l'énergie via des canevas de programmation intégrés. Pour gérer l'hétérogénéité, nous proposons une méthodologie et un support de programmation qui vise à faire communiquer les différentes entités de l’environnement ubiquitaire en utilisant le protocole SIP considéré alors comme un bus logique universel de communication. Nous avons intégré ce bus SIP dans le langage de description d’architecture DiaSpec développé par notre équipe Phoenix. Concernant la consommation d’énergie, nous proposons une méthodologie qui utilise les techniques d’offloading et de compression de données pour minimiser la consommation d'énergie des applications mobiles. Nous avons ainsi construit une stratégie d’aide à la conception au travers d’un outil qui permet de déterminer le meilleur mode d’exécution pour une tâche donnée que nous proposons d’intégrer dans le langage de description DiaSpec. / The topics of heterogeneity and energy are two fundamental considerations for pervasive computing environments. Inthis thesis, we describe our approach to manage heterogeneity and to handle energy concerns via a high-level programming framework.To manage heterogeneity, we describe a methodology and a programming support that use the SIP protocol as a universal communication bus in pervasive computing environments. Ourwork enables homogeneous communications between heterogeneous distributed entities. In doing so, we integrate the SIP communication bus into our programming framework. We rely on adeclarative language named DiaSpec to describe the architecture of pervasive applications. This description is passed to a generator for producing a Java programming framework dedicated to the application area. We leverage the generated framework with SIP adaptations to raise the abstraction level of SIP operations.We then present a classification of a wide variety of entities interms of features, capabilities and network connectors. Based on this classification, a methodology and a programming supportare described for connecting entities on the SIP communication bus. This work has been validated by applications using theSIP communication bus to coordinate widely varying entities,including serial-based sensors (RS232, 1-Wire), ZigBee devices,X10 devices, PDA, native SIP entities, and software components.Regarding the energy concerns, we describe a methodology that uses two strategies, namely computation offloading and data compression, to minimize energy cost of mobile applications.In doing so, we present an execution and transfer model for atask of a mobile application and define its five different stubs forthree program execution and data transfer modes. Based on this model and our two strategies, we construct a strategy scheme to determine the most efficient stub in terms of energy consumption.We then design the OffDeci tool, using this strategy scheme, toprovide energy feedback for the developer and to analyze thebalance between local and remote computing with consideration of data compression. Our experimental study demonstrates thefeasibility of the strategy scheme of our approach. Finally, weextend DiaSpec with declarations dedicated to manage energy concerns during the application design phase. We sketched the integration of this energy-handling declaration and OffDeci intoour high-level programming framework. This integration permitsto determine the best stub of a declared DiaSpec component interms of its energy cost.