An Experimental Evaluation of Constrained Application Protocol Performance over TCP

The Internet of Things (IoT) is the Internet augmented with diverse everyday and industrial objects, enabling a variety of services ranging from smart homes to smart cities. Because of their embedded nature, IoT nodes are typically low-power devices with many constraints, such as limited memory and computing power. They often connect to the Internet over error-prone wireless links with low or variable speed. To accommodate these characteristics, protocols specifically designed for IoT use have been designed. The Constrained Application Protocol (CoAP) is a lightweight web transfer protocol for resource manipulation. It is designed for constrained devices working in impoverished environments. By default, CoAP traffic is carried over the unreliable User Datagram Protocol (UDP). As UDP is connectionless and has little header overhead, it is well-suited for typical IoT communication consisting of short request-response exchanges. To achieve reliability on top of UDP, CoAP also implements features normally found in the transport layer. Despite the advantages, the use of CoAP over UDP may be sub-optimal in certain settings. First, some networks rate-limit or entirely block UDP traffic. Second, the default CoAP congestion control is extremely simple and unable to properly adjust its behaviour to variable network conditions, for example bursts. Finally, even IoT devices occasionally need to transfer large amounts of data, for example to perform firmware updates. For these reasons, it may prove beneficial to carry CoAP over reliable transport protocols, such as the Transmission Control Protocol (TCP). RFC 8323 specifies CoAP over stateful connections, including TCP. Currently, little research exists on CoAP over TCP performance. This thesis experimentally evaluates CoAP over TCP suitability for long-lived connections in a constrained setting, assessing factors limiting scalability and problems packet loss and high levels of traffic may cause. The experiments are performed in an emulated network, under varying levels of congestion and likelihood of errors, as well as in the presence of overly large buffers. For TCP results, both TCP New Reno and the newer TCP BBR are examined. For baseline measurements, CoAP over UDP is carried using both the default CoAP congestion control and the more advanced CoAP Simple Congestion Control/Advanced (CoCoA) congestion control. This work shows CoAP over TCP to be more efficient or at least on par with CoAP over UDP in a constrained setting when connections are long-lived. CoAP over TCP is notably more adept than CoAP over UDP at fully utilising the capacity of the link when there are no or few errors, even if the link is congested or bufferbloat is present. When the congestion level and the frequency of link errors grow high, the difference between CoAP over UDP and CoAP over TCP diminishes, yet CoAP over TCP continues to perform well, showing that in this setting CoAP over TCP is more scalable than CoAP over UDP. Finally, this thesis finds TCP BBR to be a promising congestion control candidate. It is able to outperform the older New Reno in almost all explored scenarios, most notably in the presence of bufferbloat