Network tomography by Non Negative Matrix Factorization (NNMF)

This paper presents the application of a matrix based technique to eliminate the assumption of a known routing matrix in network tomography. Network tomography is an effective means of determining network performance parameters such as delay and packet loss rate (PLR). It gives indirect inference of network characteristics using active probes or passive monitoring of packets. Most of the network tomography research unrealistically assumes that the routing matrix is known and models network tomography as an inverse problem. This motivates us to look for more appropriate methods for the inverse problem solution where the routing matrix is accommodated by the statistical ability of such methods as Non Negative Matrix Factorization (NNMF). NNMF is used to factorize a matrix into two factors (matrices). The whole process involves matrices and optimizing the residue (difference between the initial value and the current value of a cost function) to obtain the best result. We have applied NNMF to the data obtained from a laboratory test bed to perform delay tomography under various traffic conditions. The simulation results verify that NNMF performs network tomography accurately without a priori knowledge of the routing matrix.