Latency Analysis in Cloud Native Environment

In the recent years, latency is continuously assuming more and more relevance due to the demanding applications being developed. In fact, with the evolution of Industry 4.0, the arrival of autonomous driving, the increase in the videoconferencing and the growth of cloud gaming platforms, the end-to-end delay is becoming every day more a strict constraint than a marginal aspect. While it is crucial that each application provides a high-level user experience, safety-critical environments are even more demanding, as missing the requirements can have catastrophic effects. Nevertheless, besides propagation delays strictly depending on the physical distance between the two endpoints, it is necessary to consider that the underlying infrastructure, as well as higher level communication protocols, are also factors influencing network performance. Indeed, the complexity of nowadays adopted technologies (i.e. orchestration platforms such as Kubernetes, as well as network middle-boxes like load balancers and firewalls) could lead to unexpected drawbacks and even if edge computing employment is promising to reduce propagation time, this distributed paradigm may not guarantee an equal improvement in the performance perceived by applications. Is application-level latency reflecting the network-level one or is the difference in their behaviors remarkable? This thesis focuses on the analysis of what are the possible drawbacks of the cloud native environment, taking a better look at those aspects and components that could impact the performance of the end-to-end communication of a micro-service application. At first, a tool simulating a client/server application has been designed in order to gather different metrics about the communication between the two endpoints; then it has been deployed in many environment setups generating a diverse set of output data. These results showed that some scenarios worked as expected but there were some interesting behaviors that disclosed significant delay. Overall, the analysis revealed that there are not absolute problems, but depending on the specific situations some low-level tuning is typically required to better leverage the available bandwidth and achieve low latency. At the same time, small problems (e.g. a malfunctioning link corrupting some packets) turned out to have catastrophic effects, drastically decreasing the network metrics perceived by the final applications. Based on each specific scenario, the outcome has been carefully observed and some evaluations and possible solutions to the discovered problems are presented.