Introduction

Crowding is an issue whereby many Bluetooth enabled devices are in close proximity and their pairwise communication becomes so frequent that there is an increase in response time to any discovery and phone to phone communication.

In testing an early version of our protocol we found that regularly reading an ID characteristic as well as a distance estimation data characteristic (E.g. nearby ID list) would cause the continuity number to drop significantly after 4 phones were communicating.

Repeated ID fetching issue

This problem occurs most severely on earlier model phones with strange Bluetooth behaviours. This is especially true of Samsung A10, A20 and A40 devices. These devices rotate their mac address every few seconds rather than every 15 minutes. This resulted in our early protocols having to re-read identity information every few seconds. The more phones are present, the more of these superfluous identity requests were needed.

Workaround

Our current protocol uses an ephemeral ID in the advertising packet’s manufacturer data area to act as an anchor on phones that exhibit this behaviour. This has allowed us to drastically reduce the continuity error rate on tests where these phones are present.

Characteristic operation response time

All protocols use Advertising to some extent. Many non GAEN protocols will also use a read characteristic to fetch information about a remote phone. This is typically the ID characteristic, but can also include advanced distance estimation approaches too.

In our testing we found that Android phones can sometimes take several seconds, and in extremis some minutes, to complete a full read of another phone’s characteristic data. We suspect, but have not confirmed, this is an issue with how responses are returned to applications in Android’s operating system level Bluetooth stack.

Workaround

We now use write characteristics wherever possible, and never repeat any reads that we don’t have to (such as ID reads whilst the remote device has the same mac address). This has resulted in an average window size of 1-4 seconds rather than over 8 seconds using just read characteristics. It has also removed all longer than 30 second responses.

As a result our continuity error rate is much lower now than before by a considerable factor.

We therefore recommend write-with-acknowledgement wherever possible where information needs to be exchanged.