Tomofanout: a novel approach for large-scale IP traffic matrix estimation with excellent accuracy

Abstract

Traffic matrix (TM) plays an important role in many network engineering and management tasks. However, the accurate TM estimation is still a challenge because the problem is highly under-constrained. In this paper, we propose a considerably accurate approach, termed as Tomofanout, for the estimation of TM in large-scale IP network using the available link load data, routing matrix, and partial direct measurement data of TM. Firstly, we propose an edge link fanout model which defines each edge link’s fanout, i.e., each edge link’s fractions of traffic emitting from that edge link to other edge links. Secondly, benefited from the edge link fanout’s diurnal pattern and stability, we are able to compute the edge link baseline fanout to estimate the TM at the following days by multiplying it by the edge link loads at the corresponding time intervals. In such way, an initial link-to-link TM estimation result is calculated by the edge link fanout model. Further, by making the corresponding transformation to the link-to-link TM, the router-to-router TM estimation result is thus obtained. Thirdly, the solution is then refined by the basic model of the Tomography method to keep consistent with both the edge and the interior link loads for further improvement of accuracy in estimation. In particular, the expectation maximization (EM) iteration of the basic model of Tomography method is used for further refinement. As the iteration is running on, the edge link fanout model solution is gradually approaching to the final estimation result, which is compatible with both the edge and the interior link loads. Fourthly, a general algorithm is proposed for computing the edge link baseline fanout and the estimation of the TM. Finally, the Tomofanout approach is validated by simulation studies using the real data from the Abilene Network. The simulation results demonstrate that Tomofanout achieves extremely high accuracy: its spatial relative error (SRE) is less than one half of Tomogravity’s, while its temporal relative error (TRE) is less than one half of Fanout’s and is only one third of Tomogravity’s.