KMI Ranking Methodology

How KMI Rankings are determined

How does Kentik assign rankings to the ASes listed in KMI? The key assumption behind our methodology is that the volume of transited internet traffic is correlated to the amount of transited IP address space. In other words, the more transited  the IP address space, the more potential for transited internet traffic and the higher the transit score.

The volume of IP address space routed through each AS in a given geography determines its score and overall ranking. A step-by-step description of the process is provided in KMI Scoring Procedure.

BGP Data Source

The computation covered in KMI Scoring Procedure is performed daily (with greater frequency planned in the future). The BGP data we use is collected from the public RouteViews project at University of Oregon, which collects BGP data from global vantage points around the world.

Note: Future plans for KMI include expanding vantage points and visibility by augmenting the RouteViews dataset with data from our own BGP sessions with customers.

KMI Scoring Procedure

The ranking of ASes for KMI is a multi-step process, which is covered in the topics below.

Step 1: Classify AS-AS Relationships

Every AS-AS adjacency is classified as either transit or peering through a machine learning algorithm that accounts for several factors. We start with a known set of ASes that make up the default-free zone (DFZ), which is the handful of companies (e.g., NTT, Arelion, Cogent, etc.) that don’t buy transit from any other providers. With rare exceptions, AS paths exhibit a “valley-free” property that can be used to determine the transit relationships between ASes in the global routing table.

To illustrate the logic, consider the below example of ASx, which passes a route to DFZ AS1, which then passes the route to DFZ AS2:
... « DFZ AS2 « DFZ AS1 « ASx

Because DFZ ASes don’t buy transit, ASx can’t be a provider to DFZ AS1. It also can’t be a peer to DFZ AS1 without violating the valley-free rule, because relationships between DFZ ASes must be peering and two successive peering relationships are not allowed. ASx must therefore be a transit customer of DFZ AS1, and any ASes that appear to the right of ASx in an AS path should be classified as customers of ASx solely by virtue of the valley-free rule.

The above pattern can be applied until every AS-AS relationship is classified as either transit or peering. In so doing, a number of heuristics are employed to account for exceptions to the valley-free property. For example, the AS of an IXP can appear in an AS path when it is simply serving as the “glue” between two peers rather than as a direct participant in the relationship. In these cases, the IXP AS is removed from consideration in the evaluation of the relationship of the adjacent ASes.

Step 2: Score Routed Prefixes

Every routed prefix is assigned a score based on its prefix length:

• IP v4: The weight of one prefix is (1 + 24 - masklen)³ for masklen between 8 and 24 inclusive (prefixes outside this range are ignored).

• IP v6: The weight of one prefix is currently (1 + 48 - masklen)³ for masklen between 24 and 48 inclusive. However v6 prefixes between /16 and /24 get the same score, as well as prefixes between /48 and /64, which get a score of 1.

Step 3: Assign Geo and Transit Scores

Each prefix propagated across an AS-AS adjacency that has been classified as transit is assigned both a geolocation and a transit score that determine how the prefix contributes to the transit score of the provider of the adjacency:

• Geolocation: The geolocation is typically a country, which results in the prefix score contributing to the provider’s country, subcontinent, continent, and global transit scores. If the prefix is anycasted, however, it contributes only to global scores.

• Transit score: Depending on the AS hop count from the prefix's origin in the AS path, prefix scores contribute to the following types of transit scores:

  • Retail score of the origin AS as well as any single-homed upstream

  • Wholesale score of the next transit AS after the origin or single-homed upstream

  • Backbone score of the next transit AS after the wholesale AS

  • Customer base score of any AS classified as a transit provider regardless of location in the AS path

    

Step 4: Rank ASes by Transit Score

Once the geolocation and transit score contribution to each AS has been calculated for every prefix in the global routing table, Kentik has a complete set of transit scores by geolocation and score type, which means that every AS that is either originating a prefix or transiting one will have some type of transit score for at least one geography. KMI can then rank ASes by transit scores of each type for every geolocation.

Step 5: Rank ASes by Peering Score

Peer rankings apply to ASes that have peering relationships. As with transit rankings, these rankings are based on observing IP address space. The greater the amount of address space sent over a settlement-free peering session, the greater the potential for traffic to be “peered off” or sent cost-free, avoiding transit costs. So generally speaking, the more address space sent to an AS via peering, the higher the AS's peering rank will be.

ASes with peering relationships can also be ranked within a specified geography based on their overall peering base score. The peering base score for an AS is the sum of the scores for each of its peer ASes.