Back in February 2015, guest blogger Mehmet Akcin conducted a study comparing availability and latency across all of the DNS root servers, measured from hundreds of vantage points around the world.
Now that two years have passed, we’ve decided to refresh the data presented in Mehmet’s post and present an update on how the global performance of the root servers has changed. For a more detailed discussion of root servers, anycast DNS and the test methodology, see the original post here.
To replicate the original study of root server performance, we set up a variety of DNS tests. To understand availability of the root servers system as a whole, we set up a test to query for the “.” SOA record. Then, we set up a test that sends traffic to each root server to measure latency. We collected 30 days’ worth of data in order to benchmark long-term performance.
Once the tests were set up, open DNS resolvers around the world (we call them vantage points) began sending off DNS queries to the IP addresses for each root server. Because all but one of the root servers is anycasted, most of these DNS queries were routed to the nearest anycast instance for that root. With thousands of vantage points around the world testing to the root servers, we can get an understanding of how each server performs on a global scale.
Below, see how many anycast sites each root server currently has, and how those numbers have changed over the last two years. The total number of anycast sites increased by 71%, from 446 to 761 sites, with a number of root servers (D-, E-, F-, J- and K-Root) seeing significant increases in the number of anycast instances.
Availability of the Root Server System
First, we’ll look at our availability measurements to understand overall availability of the root servers — in other words, whether the DNS resolver can access the “.” SOA record at any root server in the given time period. In our test period, the DNS test sent queries every 15 minutes from roughly 2000 vantage points, 600 ASes and 50 countries around the world.
In general, we would expect overall availability to be close to 100%, as the root servers are highly redundant. Below, we break down overall availability, as well as the percentage change from 2015, by country. No country saw less than 97% availability, and all availability measurements were within roughly 3% of the same measurement taken in 2015.
Availability to the root servers has trended in a positive direction since 2015 — more than 75% of countries saw availability increase or stay the same (in the case where availability was 100% in both years).
Latency by Country
Next, we’ll look at the our latency measurements to each root server. During our test period, DNS latency tests sent queries every hour from around 300 vantage points, resulting in more than 200,000 latency measurements for every root server over the 30-day test period.
Breaking it down by geography, we found the fastest root server with the lowest mean latency in each country. Latency is closely correlated to both a root server’s response time and also the geographic distance to the server. Knowing the fastest root server in a given country is less important in a country like the US, which has many anycast sites from each root server within its borders, but for countries like Iran and Thailand, which have relatively few anycast sites, selecting to query the fastest root server can have a significantly positive impact on performance. Below, see maps for the fastest root server by country as well as the mean latency to the fastest server, as measured over the 30-day test period.
Mean latency to the fastest root server varies greatly across countries, ranging from 2 ms (Australia) to 71 ms (China). As we might expect, latency is generally higher for countries with less developed network infrastructures, like those in Asia and Latin America.
Even though great variation in latency across countries still remains since we last conducted this study in 2015, the overall trend is positive: as compared to 2015, mean latency to the fastest root server has decreased by 14 ms for each country on average.
Interestingly, the fastest root server was different from what it was in 2015 for 63% of countries. This degree of change reflects the fast pace of change in the root server landscape, where anycast sites are constantly being added. In the cases where the fastest root server had changed, the change was usually to a root server that had significantly increased its number of anycast sites since 2015.
For example, 9 more countries than in 2015 observed F-Root as the fastest — this is the biggest change across all of the root servers, so in this sense F-Root is “most improved.” F-Root has increasingly become the fastest root server in significant portions of Asia Pacific, Latin America and Eastern Europe.
This is no surprise when we consider that F-Root had 57 anycast sites in 2015, and now has 140. This total is second only to L-Root, which now has 160 sites. See the below table for a sample of countries, their fastest root servers and respective mean latencies.
F-Root is now the fastest for roughly one quarter of the countries we tested from. See the below table to see how many countries experience a given root server as the fastest.
To see all of the test data we collected for this study, explore the report snapshot here. We can also look at the Path Visualizations of extremes on both sides of the spectrum to understand why latency is so high for one country and so low for another.
If we first look at traces from Cloud Agents in Australia to the fastest root server there (F-Root), we see that the paths are short, with traffic accessing anycast sites in Australia, usually Sydney. Latency ranged from <1 to 12 ms in this particular time period.
In contrast, the traces from Cloud Agents in China to the fastest root server (also F-Root) show much longer paths with a number of issues. Traces from several Chinese agents are routed to an anycast instance in San Francisco, resulting in latencies from 152 to 181 ms, while others see roughly 60% loss in Chinese networks along the path. Only the trace from Beijing does well by global standards, which accesses a Beijing anycast site and sees 0% loss and 2 ms of latency. Degraded DNS performance in China is not unexpected, as we’ve seen in previous posts.
It’s now obvious that the factors that go into latency and performance to the root server system are varied and complex. The location of the nearest anycast site, the performance of the destination server, and the network performance and routing efficiency of the path all ultimately affect latency. On top of that, the root server system is constantly changing, as anycast sites and capacity are being added. Continuously monitor for the highest-performing root servers by setting up DNS Server tests from locations you care about — the results may prove useful for optimizing your own DNS infrastructure.