Abstract

Increasing demand for using IoT applications, such as video analytics, leverages the importance of developing an architecture to meet the requirements in terms of the latency, reliability, and energy consumption. IoT video cameras combined with the power of machine learning algorithms introduce real-time video analytics applications that can be used in diverse domains, such as security surveillance, sports, and retail stores. However, processing captured video frames using machine
learning algorithms needs resources that are beyond the capability of these IoT devices.
IoT task offloading is a new paradigm to aim IoT applications to deliver processing intensive applications to their users. IoT devices, which have limited resources by nature, offload their tasks to more powerful servers, i.e., edge/cloud servers . Nonetheless, selecting an appropriate destination for offloading the tasks is the first incoming problem for the IoT task offloading. There are some criteria which needs to be considered when it comes to IoT task offloading, for example transmission latency, queuing delay, as well as processing latency. Although edge servers have limited resources compared to cloud servers, the end-to-end latency for sending the packets to the edge servers is less than the cloud servers. On the other hand, because of the limited available resources in the edge servers, distributing the offloaded tasks between these devices is necessary to avoid overloaded servers.
Considering the above mentioned facts, in this thesis, we present load-balancing algorithms benefits from Software Defined Networking (SDN) to distribute offloaded tasks to reduce the chance of using overloaded servers and processing latency of offloaded packets of IoT video analytics applications. Taking into account the aforementioned facts, we propose a scoring metric to balance the incoming offloaded packets between edge servers. The introduced algorithm takes advantage of underlying SDN to collect information about the load of each edge server in the network. Then, the SDN controller uses the scoring metric and sorts the edge servers accordingly. The offloaded task will be directed to the edge server with the lowest processing load to avoid overloaded edge servers.
Since the number of IoT devices in the network is not predictable, increasing number of IoT devices will lead to overloaded edge servers. Hence, offloading a part of the IoT tasks to the cloud server might be a better option, even though the packets should pass through the core network. In this regard, we developed a hierarchical edge/cloud system for IoT task offloading. We modeled each of edge/cloud servers by M/M/1 queue model. By benefiting from SDN as an underlying network, the SDN calculates the processing latency and transmission latency to edge and cloud servers, and decides the best destination in terms of the minimum latency that directs the offloaded tasks to one of the desired servers.
We have conducted extensive performance evaluation to demonstrate the out-performance of the developed solutions compared with other related approaches in terms of total experienced latency and load distribution between the available servers. The results are comprehensively discussed in their related chapters to clarify the performance of the developed solution.