Recently developed techniques have been very successful in accurately estimating intra-Autonomous System (AS) traffic matrices. These techniques rely on link measurements, flow measurements, or routing-related data to infer traffic demand between every pair of ingress-egress points of an AS. They also illustrate an inherent mismatch between data needed (e.g., ingress-egress demand) and data most readily available (e.g., link measurements). This mismatch is exacerbated when we try to estimate inter-AS traffic matrices, i.e., snapshots of Internet-wide traffic behavior over coarse time scale (a week or longer) between ASs. We present a method for modeling inter-AS traffic demand that relies exclusively on publicly available/obtainable measurements. We first perform extensive Internet-wide measurement experiments to infer the ``business rationale'' of individual ASs. We then use these business profiles to characterize individual ASs, classifying them by their "utility" into ASs providing Web hosting, residential access, and business access. We rank ASs by their utilities which drive our gravity-model based approach for generating inter-AS traffic demand. In a first attempt to validate our methodology, we test our inter-AS traffic generation method on an inferred Internet AS graph and present some preliminary findings about the resulting inter-AS traffic matrices.

Papers:

• An Empirical Approach to Modeling Inter-AS Traffic Matrices [pdf]
  H. Chang, S. Jamin, Z. Mao, and W. Willinger.
  in Proceedings of ACM Internet Measurement Conference 2005, New Orleans, LA, October, 2005.

Data sets:

• Web Service Utility

  Our approach to quantifying web service utility is based on locating popular content on the Internet, as revealed by usage patterns of a popular search engine. More specifically, we obtained a list of the top 10,000 search keywords most popularly submitted to search engines in the years 2003-2004. For each keyword, we queried the Google search engine, using the Google Web API to retrieve a set of most closely matched URLs. By using these popular URLs, we infer ASs' web service utility and rank them by the inferred utility.
  • Top-10,000 keyword list
  • AS ranking for web service utility: ARIN, RIPE, APNIC
    
    

• Residential Access Utility

  To infer an AS's utility in providing residential Internet access, we estimate it by the number of P2P file sharing users of the AS. Utilizing three major P2P file sharing applications, BitTorrent, eDonkey and Gnutella, we crawled approximately 2.2 million distinct P2P IP addresses. By mapping individual IP addresses to their corresponding ASs, we obtain ASs' residential access utility and rank them by the obtained utility.
  • AS ranking for residential access utility: ARIN, RIPE, APNIC
    
    

• Business Access Utility

  Our approach to infer an AS's utility in providing business access relies on publicly available BGP routing tables to estimate the AS's bandwidth distribution. By using inferred provider-customer relationships among ASs, we approximate an AS's bandwidth distribution as the number of its downstream AS customers. We rank ASs' by the inferred bandwidth distribution volumes.
  • AS ranking for business access utility: ARIN, RIPE, APNIC