I finally decided to abandon the Ethernet piping between the airspy_adsb and the dump1090-fa and I have installed Linux natively on my Dell Latitude E6420 (dual boot).
As a result I can now pus the -e all the way to 10.0 and MLAT stays connected. Hence the timing errors were due to the network (Pi3 Ethernet port?) and not the USB like I was told. CPU usage of one core is about 73% with m12 and about 81% with m20.
Finally I was able to run two tests, with the same gain and same -e parameter, at -m 12 and -m20. Well, on my system, the performance of the -m12 is not that much better (more or less equal) than the one of -m20.
I am sure that this in not due to my computer setup. Is this is because when pushed with m20, the Airspy Mini starts to have higher noise and cancels the 20MHz advantage?
12MHz - 85 aircraft and 721 msg/s:
20MHz - 88 aircraft and 734 msg/s:
Those are a few seconds apart (edit the config, restart the services, wait for stabilization) and are in the normal fluctuation of the numbers…