A study on network performance metrics and their composition

Abstract

PurposeResearch backbone networks like GÉANT2 and the national research and education networks are used by a variety of scientists and research projects. These users and the network engineers operating the networks would like to get access to network performance metrics to optimise their use of the network and to troubleshoot performance degradations when they happen. A variety of tools for performing network measurements already exist, and the perfSONAR architecture developed in the Joint Research Activity 1 (JRA1) of GÉANT2 aims at integrating them into a coherent framework. However, a harmonised definition of the most interesting metrics and how measurements must be carried out is still lacking.Design/methodology/approachThis paper suggests some of the elementary metrics which are relevant for assessing network performance, along with an indication about how to post‐process (or “transform”, or “compose”) them in order to obtain derived summary values that can quickly and intuitively give an indication of network performance. Methods to perform the composition are presented, together with constraints which have to be taken into account to get accurate results. In particular, delay measurements are the most delicate ones to compose.FindingsThe authors carried out a series of experiments for proofing the validity of the composition of delay metrics, and briefly present some preliminary results, that confirm the applicability of the proposed methodology.Research limitations/implicationsFuture work needs to confirm the paper's findings on other data sets, possibly collected in different network locations.Practical implicationsThe practical implication of the findings is that it is possible for a network operator to accurately predict high percentiles of delays on an end‐to‐end path starting from independent delay measurements on subsequent path sections.Originality/valueThe main original contribution of this paper is the application to a real data set of a post‐processing procedure that is derived from simple statistics theory.