Search

Software architectures have continuously evolved since the advent of cloud computing, resulting in several paradigm shifts from virtualization towards microservices. Serverless computing, or Function as a Service (FaaS), has lately emerged as a new model that allows developers to focus on the development of event-driven functions, while leaving operational tasks to the serverless platforms. These functions can elastically scale to meet demand, and users are only charged for the actual execution time and consumed computing resources. Therefore, its popularity has been rising, and there are currently several serverless platforms users can choose. Such platforms include hosted platforms that are fully managed by cloud providers and open-source platforms that can be installed, for instance, on a private infrastructure. The performance of hosted platforms has been evaluated in several research works, whereas the number of studies targeting open-source platforms is still limited. Moreover, the benchmarking tools developed so far are either not actively maintained or do not target open-source platforms. Thus, the goal of the thesis is to develop a benchmarking tool to evaluate hosted and open-source serverless platforms. To this end, we leveraged FaaSTest, an open-source tool that provides a set of tests for benchmarking AWS Lambda and Azure Functions, and extended it to support several open-source serverless platforms: Knative, OpenFaaS, Kubeless, and Fission. We then benchmarked the platforms under different scenarios, including different rates of incoming traffic and computational load of the deployed functions. For the hosted platforms, the results show that AWS Lambda outperformed Azure Functions. For the open-source platforms, the evaluation results were not fully conclusive due to the challenges in configuring a large number of platform-specific features. Despite that, we observed that Knative outperformed the other solutions with a higher ratio of successful respon

Smartphones are increasingly being equipped with dedicated depth cameras that directly measure the distance to objects in a scene. These cameras enhance the 3D perception of the world and enable a wide range of immersive applications (such as 3D tele-presence and tele-surgery). However, such applications often require real-time processing of depth data (for instance, to carry out 3D reconstruction), which can be infeasible to run on a mobile device. On the other hand, the depth data may be processed on the cloud. However, this requires streaming large amounts of volumetric data. To this end, the emergence of edge computing plays a key role in providing the necessary communication and computational resources required to enable the next wave of immersive applications. The goal of this thesis is to study the feasibility of a real-time 3D tele-presence application that captures depth data from one user's smartphone and presents their volumetric 3D model to the other as a 3D avatar. We leverage a server-class edge device to overcome the limited computational capabilities of smartphones. We implement a mobile application that offloads the captured depth data to the edge server. An application on the edge server performs the 3D reconstruction that can be visualized on another smartphone or a head-mounted display. We evaluate the proposed solution in terms of the reconstruction quality and the overall performance. The experiments prove the feasibility of our solution, while achieving acceptable visual quality and reconstruction frame rate. In addition, we present several insights into designing such an application and highlight open questions for future work.

Paying by invoice has several advantages for businesses over conventional payment methods such as Debit/Credit cards. An invoice not only allows a buyer to make the purchase on credit but also contains the tax information for each purchased item. Businesses need to save this information in their financial records and report it to the authorities. The core challenge in an invoice-based payment method is the ability to make an accurate credit decision for a given purchase. Such a credit decision requires information about the buying company such as their credit rating. The company information is gathered in real time from different third-party sources. In this context, Enterpay Oy provides an invoiced-based B2B payment solution and is growing its payment service to European countries. In order to support this expansion, Enterpay needs to develop new capabilities such as the ability to detect fraudulent purchases. These new features require the application architecture to be flexible in terms of technology. For example, different components of the service should be built with the most suited programming language, libraries, and frameworks. The goal of this thesis is to enable efficient scaling and high availability for Enterpay's payment service. Thus, we have migrated from a monolithic a microservice-based architecture. This transition allows us to choose the best suited technology for the business case of the given microservice. We extracted various modules from the original monolithic application, which have different scalability criteria. We built these modules as Docker containers, which run as independent microservices. We used Kubernetes as the container orchestration framework and deployed the microservice in Amazon Web Services (AWS). Finally, we conducted experiments to measure the performance of the service with the new architecture. We found that this architecture not only scales faster but also recovers from instance failures quicker than the previous solu

Access management is a key aspect of providing secure services and applications in information technology. Ensuring secure access is particularly challenging in a cloud environment wherein resources are scaled dynamically. In fact keeping track of dynamic cloud instances and administering access to them requires careful coordination and mechanisms to ensure reliable operations. PrivX is a commercial offering from SSH Communications and Security Oyj that automatically scans and keeps track of the cloud instances and manages access to them. PrivX is currently built on the microservices approach, wherein the application is structured as a collection of loosely coupled services. However, PrivX requires external modules and with specific capabilities to ensure high availability. Moreover, complex scripts are required to monitor the whole system. The goal of this thesis is to make PrivX highly-available and scalable by using a container orchestration framework. To this end, we first conduct a detailed study of mostly widely used container orchestration frameworks: Kubernetes, Docker Swarm and Nomad. We then select Kubernetes based on a feature evaluation relevant to the considered scenario. We package the individual components of PrivX, including its database, into Docker containers and deploy them on a Kubernetes cluster. We also build a prototype system to demonstrate how microservices can be managed on a Kubernetes cluster. Additionally, an auto scaling tool is created to scale specific services based on predefined rules. Finally, we evaluate the service recovery time for each of the services in PrivX, both in the RPM deployment model and the prototype Kubernetes deployment model. We find that there is no significant difference in service recovery time between the two models. However, Kubernetes ensured high availability of the services. We find that Kubernetes is the preferred mode for deploying PrivX and it makes PrivX highly available and scalable.

Recent advancements in virtualization and software architectures have led to the birth of the new paradigm of serverless computing. Serverless computing, also known as function-as-a-service, allows developers to deploy functions as computing units without worrying about the underlying infrastructure. Moreover, no resources are allocated or billed until a function is invoked. Thus, the major benefits of serverless computing are reduced developer concern about infrastructure, reduced time to market and lower cost. Currently, serverless computing is generally available through various public cloud service providers. However, there are certain bottlenecks on public cloud platforms, such as vendor lock-in, computation restrictions and regulatory restrictions. Thus, there is a growing interest to implement serverless computing on a private infrastructure. One of the preferred ways of implementing serverless computing is through the use of containers. A container-based solution allows to utilize features of existing orchestration frameworks, such as Kubernetes. This thesis discusses the implementation of serverless computing on Kubernetes. To this end, we carry out a feature evaluation of four open source serverless computing frameworks, namely Kubeless, OpenFaaS, Fission and OpenWhisk. Based on predefined criteria, we select Kubeless, Fission and OpenFaaS for further evaluation. First, we describe the developer experience on each framework. Next, we compare three different modes in which OpenFaaS functions are executed: HTTP, serializing and streaming. We evaluate the response time of function invocation and ease of monitoring and management of logs. We find that HTTP mode is the preferred mode for OpenFaaS. Finally, we evaluate the performance of the considered frameworks under different workloads. We find that Kubeless has the best performance among the three frameworks, both in terms of response time and the ratio of successful responses.

Network Functions Virtualisation (NFV) involves the implementation of network functions, for example firewalls and routers, as software applications that can run on general-purpose servers. In present-day networks, each network function is typically implemented on dedicated and proprietary hardware. By utilising virtualisation technologies, NFV enables network functions to be deployed on cloud computing infrastructure in data centers. This thesis discusses the application of NFV to the Evolved Packet Core (EPC) in Long Term Evolution (LTE) networks; specifically to the Mobility Management Entity (MME), a control plane entity in the EPC. With the convergence of cloud computing and mobile networks, conventional architectures of network elements need to be re-designed in order to fully harness benefits such as scalability and elasticity. To this end, we design and implement a distributed MME with a three-tier architecture common to web applications. We highlight design considerations for moving MME functionality to the cloud and compare our new distributed design to that of a standalone MME. We deploy and test the distributed MME on two separate OpenStack clouds. Our results indicate that the benefits of scalability and resilience can outweigh the marginal increase in latency for EPC procedures. We find that the latency is dependent on the actual placement of MME components within the data center. Also, we believe that extensions to the OpenStack platform are required before it can meet performance and availability requirements for telecommunication applications.