Data-Driven End-to-End Delay Violation Probability Prediction with Extreme Value Mixture Models

Abstract

With the advent of edge computing, there is increasing interest in wireless latency-critical services. Such applications require the end-to-end delay of the network infrastructure (communication and computation) to be less than a target delay with a certain probability, e.g., 10-2-10-5. To deal with this guarantee level, the first step is to predict the transient delay violation probability (DVP) of the packets traversing the network. The guarantee level puts a threshold on the tail of the end-to-end delay distribution; thus, it makes data-driven DVP prediction a challenging task. We propose to use the extreme value mixture model in the mixture density network (MDN) method for this task. We implemented it in a multi-hop queuing-theoretic system to predict the DVP of each packet from the network state variables. This work is a first step toward utilizing the DVP predictions, possibly in the resource allocation scheme or queuing discipline. Numerically, we show that our proposed approach outperforms state-of-the-art Gaussian mixture model-based predictors by orders of magnitude, in particular for scenarios with guarantee levels above 10−2.