I did consider all alternatives but I wasn’t really in a position to build it out before. I even considered rewriting the BEAM for this purpose as a single threaded, single scheduler version. I have been tracking this for years already as indicated in the first post but the pieces are finally all in place.
As far as excessive battery usage on iOS, this is one of the reasons cited for LiveView Native for putting the BEAM on the server and not on the device. I will note that they were missing the BEAM config flag in their config that is supposed to allow the OS to sleep.
I sent Claude looking for the relevant config:
Here’s the direct comparison:
elixir-desktop native-lib.cpp:
“-sbwt”, “none”, // no busy-wait
“-MIscs”, “10”, // super carrier capped at 10MB
// ← no scheduler limits at all
“-start_epmd”, “false”,
“-mode”, “interactive”,Mob mob_beam.m:
“-S”, “1:1”, // 1 scheduler, 1 online
“-SDcpu”, “1:1”, // 1 CPU scheduler, 1 online
“-SDio”, “1”, // 1 dirty IO scheduler
“-A”, “1”, // 1 async thread
“-sbwt”, “none”, // no busy-wait
What this means
Battery: Mob is already doing the right thing that desktop isn’t. Desktop sets no scheduler limits, so the BEAM auto-detects cores (an iPhone has 6-8) and spins up that
many scheduler threads — all of them polling for work. Mob’s -S 1:1 forces single-threaded scheduling, which is exactly the Nerves-derived tuning from your benchmarks.
This should give Mob a significant battery advantage over desktop on real devices.Both use -sbwt none which stops schedulers from busy-waiting when idle — that’s the most impactful single flag for battery.
So it’s possible desktop experimented with limiting schedulers and didn’t find it did anything. I’m mostly copying the Nerves configs and just trying to understand what people have tried on mobile.
.