Sliced WANs for Data-Intensive Science: Deployment Experiences and Performance Analysis
Edgard C. Pontes, Vitor F. Zanotelli, Magnos Martinello, Jordi Ros-Giralt, Everson Scherrer Borges, Moisés R. N. Ribeiro, H. B. Newman · 2024
The task of transferring massive data sets in Data-Intensive Science (DIS) systems, such as those generated from high energy experiments at CERN in Switzerland and France, and at the Sirius synchrotron light source in Brazil, often rely on physical WAN infrastructure for network connectivity that is provided by various National Research and Education Networks (NRENs), including ESnet, Géant, Internet2, RNP, among others. Sliced WANs bring a new paradigm for infrastructure yet to be exploited by DIS, but a realistic study of these particular systems poses a significant challenge due to their complexity, scale, and the number of factors affecting the data transport. In this paper, we address some of these challenges by deploying and evaluating a virtual infrastructure for data transport within a representative national-scale WAN. Our approach here encompasses two main aspects: i) Evaluating the performance of TCP congestion control algorithms (BBR versus Cubic) when only a single path is available for the data transfer; and ii) Assessing the performance of flow completion times (related to the management of bandwidth allocation) for sets of interdependent transfers in an environment provided by a network slice.