Mapping prediction with recurrent neural networks for future LISP enabled networks

Abstract

Locator/identifier separation paradigm (LISP) is an emerging Internet architecture evolution trend that decouples the identifier and location of an entity attached to the Internet. Due to its flexibility, LISP has seen its application in various fields such as mobile edge computing, and V2X networks. However, LISP relies on a DNS-like mapping system to associate identifiers and locations before connection establishment. Such a procedure incurs an extra latency overhead and thus hinders the adoption of LISP in delay-sensitive use cases. In this paper, we propose a novel RNN-based mapping prediction scheme to boost the performance of the LISP mapping resolution, by modeling the mapping procedure as a time series prediction problem. The key idea is to predict the mapping data regarding services to be utilized by users in edge networks administered by xTRs and proactively cache the mapping information within xTRs in advance. We compare our approach with several baseline methods, and the experiment results show a 30.02% performance gain in LISP cache hit ratio and 55.6% delay reduction compared with the case without mapping prediction scheme. This work preliminarily proves the potential of the approach in promoting low-latency LISP-based use cases.