Traffic Matrix Modelling Sample Clauses

Traffic Matrix Modelling. With a view to potentially extending perfSONAR with traffic matrix (TM) functionality, research is conducted to provide good models for TMs, with potential applications for traffic demands prediction, anomaly detection, and traffic engineering. Traffic matrices offer a way to express the volume of data exchanged between ingress and egress nodes in a network [ACRUW06]. Although TMs can be expressed at different levels of temporal (from 5 minutes to hours, days, months or years) and spatial aggregation (origin-destination machine, OD prefix, OD router, OD PoP or OD AS), the investigation will focus on cases of interest to GÉANT, and the nodes will be either routers or the PoPs present at each NREN. While there is some literature on the analysis of TM, few efforts have been made in the field of multi-domain TMs. In the past, a lot of effort has been devoted to the problem of measuring TMs. Although there are tools that can provide TMs directly (e.g. NetFlow), they do not scale well (see Enhanced Flow-Level Measurements on page 37), and measurement campaigns are difficult to set up (see [UQLB06] for a detailed description of one of such measurement campaigns, performed in GN2). On the other hand, SNMP data (describing the volume of traffic sent or received in each link) is easy to collect, and almost ubiquitous, as most networking equipment supports the MIB-II. However, SNMP data only provides link load measurements, not TM measurements, i.e., SNMP is able to compute the aggregate traffic transported by the link, but unable to distinguish the specific origin- destination pairs of each flow. The resulting problem of inferring the TM from link measurements is a classic under constrained, linear-inverse problem which needs some sort of side information, such as a Poisson model [Var96] or a Gaussian model for the TM entries, or the use of a previous model based on the gravity relationships among nodes (the so-called “gravity model” [ZRDG03]), among others. The challenge of finding a good model for TMs is application-dependent, and therefore quite open. The criterion used for defining a TM model is sparsity, i.e., the model should have a small number of parameters. In the case of TMs, a network of N nodes has a TM of N2 entries (N origins x N destinations), and for typical sizes of N in the range of hundreds or even thousands, the TM can become too large to manage easily (for example, when storing TMs, predicting traffic or detecting anomalies). This sparse model w...