Keep your TURN servers close
When Amazon AWS opened the new region in London, we quickly spun up new TURN servers there. What impact did this have on the ICE candidate gathering times?
As I explained in an earlier blog post, the ICE candidate gathering time can be a useful proxy metric for the call setup time. It is easier to look at than the call setup time because it only depends on the distance between a single client and the TURN server and not on the locations of both clients. It obviously depends heavily on where your TURN servers are located in relation to your client, which got me into an argument about aggregation this earlier this week. It also depends on whether the client is configured with only TURN/UDP or also TURN/TCP and TURN/TLS which, no kidding, take longer because of handshakes.
When we deployed a TURN server to the new AWS region in London I wanted to check our hypothesis that we should spin up turn everywhere to lower latency. Bear with me…
For November 2016 and the first half of December, the weekly median was fairly stable, around 225ms with larger daily variations. The clients from the UK were connecting to the Frankfurt datacenters during that period.
The gathering time dropped slightly to 218ms after deploying TURN servers in London on December 19th. At the same time usage patterns changed significantly due to the holiday season, which affects our metrics and our ability to interpret them.
During the first week of January 2017 the gathering time dropped to 198ms. That was a good sign but usage was not yet back to normal.
January 9th was the first day which could be considered normal again in terms of usage and there was a significant drop to 157ms. Tuesday continued with 155ms, Wednesday 165ms, Thursday 149ms and Friday 158ms. That is an average of 155ms and significantly lower than the 225ms we have seen before.
And just because we deployed a new TURN server which is a matter of minutes. Its a very low-effort and high-impact way to improve the quality of your service. And it seems that our users noticed, appear.in saw quite some growth from the UK since we deployed servers there. Of course that might be unrelated but it is nice to pretend that there is a correlation :-)
