Context-aware task scheduling in distributed computing systems

These days, the popularity of technologies such as machine learning, augmented reality, and big data analytics is growing dramatically. This leads to a higher demand of computational power not only for IT professionals but also for ordinary device users who benefit from new applications. At the same time, the computational performance of end-user devices increases to meet the demands of these resource-hungry applications. As a result, there is a coexistence of a huge demand of computational power on the one side and a large pool of computational resources on the other side. Bringing these two sides together is the idea of computational resource sharing systems which allow applications to forward computationally intensive workload to remote resources. This technique is often used in cloud computing where customers can rent computational power. However, we argue that not only cloud resources can be used as offloading targets. Rather, idle CPU cycles from end-user administered devices at the edge of the network can be spontaneously leveraged as well. Edge devices, however, are not only heterogeneous in their hardware and software capabilities, they also do not provide any guarantees in terms of reliability or performance. Does it mean that either the applications that require further guarantees or the unpredictable resources need to be excluded from such a sharing system? In this thesis, we propose a solution to this problem by introducing the Tasklet system, our approach for a computational resource sharing system. The Tasklet system supports computation offloading to arbitrary types of devices, including stable cloud instances as well as unpredictable end-user owned edge resources. Therefore, the Tasklet system is structured into multiple layers. The lowest layer is a best-effort resource sharing system which provides lightweight task scheduling and execution. Here, best-effort means that in case of a failure, the task execution is dropped and that tasks are allocated to resources randomly. To provide execution guarantees such as a reliable or timely execution, we add a Quality of Computation (QoC) layer on top of the best-effort execution layer. The QoC layer enforces the guarantees for applications by using a context-aware task scheduler which monitors the available resources in the computing environment and performs the matchmaking between resources and tasks based on the current state of the system. As edge resources are controlled by individuals, we consider the fact that these users need to be able to decide with whom they want to share their resources and for which price. Thus, we add a social layer on top of the system that allows users to establish friendship connections which can then be leveraged for social-aware task allocation and accounting of shared computation