DNS Optimizations with Dnsmasq

DNS Optimizations with Dnsmasq

We currently fetch milions of feeds and we need to do that in a very timely fashion. Since the beginning of 2014, it took about 600ms to fetch a feed on average. The fetching time includes DNS resolution, the establishment of the HTTP connection, the HTTP transfer, and the decoding of the answer (if it was gzipped for example). This is by far what takes the longest wall time.

Luckily, we use asynchronous IO, which means, that we never really wait idly for the data to come: we usually start processing other feeds. However, this significantly increases the memory consumption, and eventually has a toll on performance. For this reason, it’s important that we optimize our requests as much as possible. This starts with DNS.

Dnsmasq locally

Dnsmasq is a lightweight, easy to configure DNS forwarder…

We currently have 20 servers in charge of polling resources on the web. They perform the bulk of our DNS requests. They’re now all running a Dnsmasq instance. Dnsmasq is only efficient if it stays small, so we limited the size of the cache to 1024 entries.

To test the impact of Dnsmasq, we built a quick benchmark script. We extracted 10k urls from our database, and extracted their domains. We kept the buggy domains in them. We then used this simple node.js script which takes a domain and resolves it with dns.resolve before handling the next one.

We changed the content of /etc/resolv.conf to include only the server which will be used by Node to resolve the requests. We tested with 3 different ones and here are the results:

  • One of Google’s public DNS ( Min:7 Max:4097 Median:8 Average:33.66747
  • One of OpenDNS’s public servers ( Min:29 Max:6030 Median:30 Average:100.3912
  • One of Linode’s local resolvers ( Min:49 Max:67166 Median:52 Average:183.1860

The point here is not to compare these servers, but to see the impact of using a cache locally.


With dnsmasq installed and as the first line of /etc/resolv.conf, we get the following results:

  • Google: Min:0 Max:3028 Median:8 Average:18.41814
  • OpenDNS: Min:0 Max:8768 Median:29 Average:66.38854
  • Linode: Min:0 Max:41248 Median:49 Average:110.6278

The resolution time went down to 0 for 3605 requests, affecting the average pretty significantly (between 30% and 50%).

In practice, we get a hit rate of about 50% in our production environment, which is quite high and saves us a significant wall time. A 50% hit rate on our local caches is also a 50% miss rate, we’ll see next how we can improve further our DNS resolution performance…

Liked this post? Read the archive or

On the same topic, check dns optimizations with unbound, moving to prosody.im and a new pubsubhubbub endpoint.

Previously, on the Superfeedr blog: Medium supports PubSubHubbub.