🔧 Debugging Mari's Performance - From 50% to 98% Packet Delivery

You can only optimize what you can measure. And measurement is crucial when dealing with scalable systems (100+ real-time nodes).

For a few weeks I was puzzled by the unstable success ratio of Mari (our new link layer for micro-robots), especially the performance degradation during over-the-air updates of large robot fleets.

This week, I finally took the time to dig deep. That meant:

Instrumenting both the link layer and the host library to collect and show relevant metrics in real time, including packet delivery ratio (PDR) for both the radio and serial links.
Observing execution patterns using a logic analyzer, toggling pins in carefully chosen code paths.

This led to a precise diagnosis: the radio was doing fine, but the serial link was suffering from concurrency issues and lost interrupts.

The Solution

After a set of adjustments:

Re-prioritizing interrupts
Adding a queue
Making a task non-blocking

The improvement was drastic. Average PDR jumped from ~50% with 10 nodes, to 98% on the radio link and 92% on the serial connection with 100 devices.

There’s still room for refinement, but the current release already provides serious value for users.

Learn More

➡️ We are building open-source tools for robotic swarms:

Check out the project page at https://openswarm.eu/
Our code at https://github.com/dotbots

Before and after

PDR before optimization: ~50% success rate with only 10 nodes

PDR after optimization: 98% radio link and 92% serial connection with 100 devices