How well are we doing with the adoption of Route Origin Validation (ROV) in the interdomain routing space?
There have been many workshops and training sessions, and much in the way of counting the generation of Resource Public Key Infrastructure (RPKI) certificates and Route Origin Attestations (ROAs) in recent months. The data published by the US National Institute of Standards and Technology (NIST) in its RPKI monitor is a good example. Around 20% of the announced prefix/origin AS pairs have an associated valid ROA. This is a good example of production measurements that track the adoption of technology through the measurement of the infrastructure itself.
However, there is another aspect of the measurement of this technology, and that is as a user measurement. How many users can no longer reach a destination if the only available ROAs mark the destination announcement as invalid?
Ideally, if this technology was fully deployed, then the answer to such a question would be ‘all users’. But what do we mean by ‘fully deployed’ in the routing system? This might sound like an odd question, but it’s useful to remember that the topology of the interdomain routing system is not a fully interconnected mesh. It is far more like a hierarchical topology where the overall majority of networks are stub networks that have a single connection to an upstream transit network.
Using the vantage point of the Route-Views Route Collector, there are currently some 58,000 such stub networks and 11,000 transit networks in the IPv4 topology and 16,000 stub networks and 4,000 transit networks in IPv6.
To prevent users from reaching a destination it is only necessary at a global level to remove the route from the transit networks. This is to some extent a simplification, but it does contain an essential truth: to remove a route from the Internet at large, it’s only necessary to remove it from the ‘default-free zone’ on the routing system, as well as clearing it out from local peering sessions.
This removal of an invalid route from the default-free zone of the Internet gives us a method to measure the effectiveness of uptake of Invalid ROA filtering through the impact of reachability. In this measurement, we have set up an announcement of a route into the default-free zone, and we periodically change the ROV status of the route from valid to invalid and back by revoking the previous ROA and issuing a new ROA at each phase transition.
Phase changes in ROAs
How long should we maintain a ROA state? Is the scale of this system sensitive to states that last only a few seconds, or does the state need to be held for hours, or even days, in order for all validating entities to detect the ROA change?
At APNIC Labs, set up a delegated RPKI system for this measurement. Figure 1 shows the average re-query interval for each RPKI client (as determined by a unique IP address) over a 90-day measurement window.
There are three peaks in terms of re-query intervals (Figure 2).
We can now identify these peaks as 2 minutes, 10 minutes and one-hour re-query intervals. By taking a cumulative distribution we can determine the distribution of clients across re-query intervals.
Some 10% of clients have an average re-query interval of 120 seconds or less. This rises to 25% of clients who use an average re-query interval of 630 seconds. Some 66% of clients use an average re-query interval of 60 minutes or less. Some 70% of clients have an average re-query interval of 70 minutes or less. There is no further clustering of re-query intervals for longer periods after 70 minutes.
The implication of these numbers is that if we use a phase change of ROA validity in terms of 12 hours or more, we are confident that at least 90% of all RPKI clients will pick up the state change.
For the purposes of this measurement exercise, we use a mix of 12-hour and 36-hour intervals and flip a route advertisement between ROA validation states of valid and invalid in this period.
The weekly schedule of ROA validity state is shown in Figure 5.
The route has an invalid ROV status for half of each week, spread into two 36-hour intervals and one 12-hour interval.
We can confirm a change of routing state corresponding to this schedule of changing the ROV validity status of this route by looking at any of the BGP update monitors.
- If a network’s BGP speakers are performing ROV filtering, then when the ROA changes the route to refer to an incorrect Origin AS, the route will then be marked as being invalid.
- If the network is performing filtering of routes where the route is invalid (as distinct from unknown) then this network will not only withdraw the route from its own network but withdraw any announcements of this route from its routing neighbours.
- Similarly, when the ROA changes the route validation state back to referring to the actual Origin AS (revoking the previous ROA) the Route Origin Validation state will switch to valid, and if the network had previously filtered the route it will then announce the route both internally and to its routing neighbours.
These route changes and their propagation through the eBGP interdomain space should be visible to the route collectors. The route update profile for this prefix as seen by the RIPE ROA collectors is shown in Figure 6, and the routing activity peaks match the scheduled changes of ROA states shown in Figure 5. This BGP monitor data suggests that some networks are performing filtering of invalid routes.
We then use the measurement platform to have end-user hosts retrieve a web plot from a server that lies behind this route. By measuring the extent to which reachability of the routed address changes between periods when the ROV state is valid and invalid we can measure the extent to which the Internet users are located behind routing systems that filter invalid routes.
It should be noted that being unable to reach an ROV-invalid does not necessarily mean that the local network is performing ROV and filtering invalid routes. It could be the network’s upstream transit provider that is performing this action or any transit provider on the path between the network where the test user is located and the point of route injection.
The results of this test are dependent on the location of the route injection. For example, if the test system sat directly behind a transit provider that performed filtering of invalid routes then it would appear as if the entire Internet was performing filtering of invalid routes! To mitigate this potential measurement bias we are currently using three diverse locations to anycast the route origination. This diversity is not only geographic but diverse in terms of the transit service providers used for the test.
The Internet-wide per-second profile of reachability of this prefix for an entire week is shown in Figure 7. The data points in green were collected during periods in the week when the route status was valid. Red denotes those periods when the route was invalid.
There is certainly some impact on reachability when the ROV state is invalid. In this experiment, the average fetch rate in the time when the route is valid is an average of 95%, while the periods when the route is invalid the reachability levels drop to 80 to 83%. This infers that filtering invalid routes in the Internet currently is effective for some 10% of the total user population.
It’s interesting to observe in this data that the fall in reachability levels when the route status changes from valid to invalid is far slower than the corresponding rise in reachability in the transition from invalid to valid.
Figure 8 shows the valid to invalid transition in greater detail. The major transition occurs 30 minutes after the change in ROA status.
The invalid to valid transition is a lot faster, as shown in Figure 9.
Here, the transition takes a little under 300 seconds to take effect.
Why are the transitions different?
The test route is advertised into the routing system in three distinct locations. While some transit networks may react quickly, other transit networks may be using slower timers.
They may be performing a scan of the RPKI publication points every 10 minutes or even every hour, so it may be some time before the ROA state change is noticed. Then, in the case of the valid to invalid transition, it is the time of the slowest transit paths to be withdrawn that determines the overall transition, as any initial withdrawals may simply cause the traffic to pick an alternate path. It is only when all alternates are withdrawn that the route becomes unusable. That means that the transition time is largely determined by the slowest networks to react, which from this data looks to be around 30 minutes.
The opposite transition is largely governed by the fastest networks to react. As more paths become available the traffic may shift to other paths, but the transition time is determined by the fastest reaction times.
Measuring the user impact of invalid route filtering
The aim of this measurement is not to measure which networks are performing invalid route filtering, but to measure the user impact of this filtering. That is to say, we are interested in understanding which invalid route filtering is effective in preventing users from reaching destinations announced through invalid routes and the trend over time of this measurement. The initial set of results for this measurement are shown in Figure 10.
Through June 2020, RPKI filtering impacts some 17% of the Internet’s user population. As already noted, it is clear that this is likely the result of the uptake of filtering by a small set of transit providers rather than a larger deployment in the collection of stub networks.
We can break this down into a view per-geographic region and further down into a per-economy view. The results, shown in Figure 11, is somewhat surprising, in that invalid route filtering is most prevalent for users in Africa.
It’s surprising because none of the RPKI beacon points are located in this region so it’s possible that this figure reflects the selection of transit providers in this part of the world. Where a transit service performs filtering the customer networks of this service see the route withdrawal and announcement in line with the ROA status schedule shown in Figure 5.
In the United States, AT&T (AS7018) was an early adopter of invalid route filtering, and the measurement system also sees users of this network following the reachability dictated by the route validity status schedule (see results to date).
It appears that NTT Communications (AS2914) also performs invalid route filtering, although the measurement through June 2020 shows some issues with this in the period 7-17 June in the US (see results to date).
There are a lot of moving parts in RPKI — ROA generation and publication and use of ROAs to perform validating and invalid route filtering — so perhaps it’s unsurprising to see periods where the total system is not operating as intended.
In some economies, we see some local providers perform filtering while others do not.
- Bhutan is a good case in point where TashiCell appears to perform invalid route filtering directly while DrukNet and Bhutan Telecom do not. See RPKI validation for Bhutan.
- In Sweden, Telenor and Telia appear to perform invalid route filtering, while Tele2, in Sweden, is not. See RPKI validation for Sweden.
- In France, Proxad (AS12322) (sold to consumers under the name ‘Free’) was an early adopter of IPv6 and it is also an adopter of invalid route filtering. See RPKI validation for France.
We will continue this daily process of gathering measurements and publishing global, regional and per-network measurements over the coming months and years to assemble a comprehensive picture of the growing adoption of invalid route filtering as seen from the user perspective — see RPKI validation measurement results for your economy/region.
The views expressed by the authors of this blog are their own and do not necessarily reflect the views of APNIC. Please note a Code of Conduct applies to this blog.