Networked Latency Sensitive Applications - Performance Issues between Cloud and Edge

Abstract: The increasing demand for industrial automation has motivated the development of applications with strict latency requirements, namely, latency-sensitive applications. Such latency requirements can be satisfied by offloading computationally intensive tasks to powerful computing devices over a network at the cost of additional communication latency. Two major computing paradigms are considered for this: (i) cloud computing and (ii) edge computing. Cloud computing provides computation at remote datacenters, at the cost of longer communication latency. Edge computing aims at reducing communication latency by bringing computation closer to the users.  This doctoral dissertation mainly investigates relevant issues regarding communication latency trade-offs between the aforementioned paradigms in the context of latency-sensitive applications.This work advances the state of the art with three major contributions. First, we design a suite of scheduling algorithms which are performed on an edge device interposed between a co-located sensor network and remote applications running in cloud datacenters. These algorithms guarantee the fulfillment of latency-sensitive applications' requirements while maximizing the battery life of sensing devices.  Second, we estimate under what conditions latency-sensitive applications can be executed in cloud environments. From a broader perspective, we quantify round-trip times needed to access cloud datacenters all around the world. From a narrower perspective, we collect latency measurements to cloud datacenters in metropolitan areas where over 70% of the world's population lives. This Internet-wide large-scale measurements campaign allows us to draw statistically relevant conclusions concerning the readiness of the cloud environments to host latency-sensitive applications. Finally, we devise a method to quantify latency improvements that hypothetical edge server deployments could bring to users within a network. This is achieved with a thorough analysis of round-trip times and paths characterization resulting in the design of novel edge server placement algorithms. We show trade-offs between number of edge servers deployed and latency improvements experienced by users.This dissertation contributes to the understanding of the communication latency in terms of temporal and spacial distributions, its sources and implications on latency-sensitive applications.