Performance of the DNS Transport Layer

In our first test we wanted to investigate how long it takes to service a single client. Therefore, we measured an average Round Trip Time (RTT), which is the time between sending a DNS request and receiving a response. To make the results as accurate as possible, we did not use any physical network connection. Instead, we placed both the client and the Apache server running our redirector on the same machine. Also, to avoid any concurrency-related inaccuracies, we configured Apache to use only one request-processing thread.

RTTs were measured in the following way. A client sequentially issued 2000 DNS requests that regarded the same, fictitious domain. We measured the time of calling the ``res_send'' function, implemented inside the system resolver library, and calculated an average of all the results. This test was run using TCP-based and UDP-based queries.

Additionally, to measure the overhead introduced by different policies, we performed the above two tests for both the static name-to-address policy, and the round-robin policy.

Finally, to discover the correspondence between the number of IP addresses returned in a single DNS response and the actual RTT, we ran all the tests on four separate redirector configurations, each responding with a different number of addresses. The results of our tests are summarized in Figure 5.1.

Figure 5.1: Round Trip Times for simple DNS queries

As can be observed, there is a significant difference in RTTs between queries using TCP and UDP: retrieving a DNS response with TCP takes almost 80% more time than doing that with UDP. This difference is probably caused by the cost of establishing a TCP connection.

The overhead introduced by the round-robin policy turned out to be negligible in comparison to total execution time. Similarly, the increase in number of IP addresses returned in a single DNS response results in insignificant growth of the overall RTT.