Prometheus in Practice - Communication-Efficient Monitoring on Distributed Resource Constrained Systems

Abstract

Continuous Distributed Monitoring (CDM) aims to reduce the communication over head associated with the monitoring of a distributed system. This thesis investigates forecast-based monitoring techniques as a practical solution to the CDM problem by applying them in a Prometheus-based monitoring framework. The investigation resulted in a proxy-based framework that interfaces with Prometheus and implements two prediction models, Simple Approximation and Piecewise Linear Approximation. Evaluation was conducted in two phases: one simulation-based phase and one hardware-based phase, in which the framework was ported to a physical test bench of Raspberry Pis. The final framework was evaluated in terms of communication ratio, mean absolute error, update latency, as well as CPU and memory usage. On average, the final results indicated a 77.4% decrease in transmitted network packets while keeping the mean error at 1.13%.