Skip to content

ASB-V36

Choose a tag to compare

View all tags
@krilikd krilikd released this 10 Apr 20:23
· 3572 commits to main since this release
36
9a4360c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

AutoSystemBoost -- Changelog

V36 -- Long Gaming + Deeper Battery + Honest Status

V35 fixed the thermometer and calmed balanced down.
V36 rebuilds the performance profile around sustainable 1-2 hour gaming sessions, pushes battery into deeper energy savings, exposes honest telemetry in status, and fixes a silent install error that was hiding since V27.

🎮 Performance Profile -- Built for Long Gaming

V35 performance was optimized for peak 60-second bursts. Logs showed this caused early vendor ceiling collapse around 30s → low actual sustained performance over a 1-2 hour session.

V36 rebuilds performance around a sustainable ceiling philosophy: don't chase peaks, ride a stable plateau at 40-55C shell temperature that the vendor clamp never bites.

Parameter V35 V36 Why
CPU_CAP_BIG 3,456,000 3,264,000 Soft cap -200 MHz keeps vendor clamp quiet
CPU_CAP_LITTLE 3,072,000 2,995,200 Matched ratio
CPU_MIN_LITTLE 1,497,600 1,324,800 Lower floor = heat bleed between frames
WALT_ED_BOOST 55 → 40 35 Less thermal spike, frames still hit 144fps
WALT_TOPAPP_WEIGHT 190 → 170 160 Calmer boost, game still prioritized
WALT_BUSY_HYST 12 ms 8 ms Faster downscale when frame is done
SCHED_RATE 800 1,000 Fewer unnecessary boost events mid-frame
SCHED_UP_RATE 300 400 Slightly slower up-scaling
SCHED_HISPEED_LOAD 55 60 Higher bar for hispeed entry
GPU_IDLE_TIMER 48 64 Less aggressive GPU nap transitions
GPU_MAX_PCT 100 92 Reserve 8% GPU headroom
GPU_THERMAL_PWRLEVEL 1 2 One step back from absolute max
VM_DIRTY_EXPIRE 1000 1500 Less kswapd pressure mid-session
UCL_FG_MAX 100 95 Foreground gets 95%, top-app 100%

Philosophy: Don't cut sustained after the fire — don't start the fire. Lower opening punch → vendor clamp stays asleep → higher actual sustained frequencies over 1-2 hours of COD Mobile at 144fps.

Expected result: shell temperature 40-55C instead of 60-70C spikes, fewer SUSTAINED entries, stable frame pacing over 60+ minute sessions.

🔋 Battery Profile -- Deeper Energy Saving

V35 battery was already good for sleep (clean 8h nights). V36 goes further on screen-off and mixed-day energy drain.

Parameter V35 V36 Why
CPU_MIN_LITTLE 384,000 307,200 Absolute minimum idle floor
CPU_MIN_BIG 768,000 614,400 Lower big-cluster floor
CPU_MAX_BIG 1,132,800 998,400 Cap big-cluster ceiling even harder
CPU_CAP_LITTLE 729,600 614,400 Soft cap matches min floor
CPU_CAP_BIG 1,075,200 921,600 No unnecessary big-cluster activity
WALT_TOPAPP_WEIGHT 60 40 Even foreground gets less priority
WALT_BOOST_MIN_UTIL 150 200 Harder to justify frequency boost
SCHED_RATE 12,000 16,000 Slower scheduler = fewer wake events
UCL_FG_MAX 10 8 Foreground apps capped at 8% utilization
UCL_TOP_MAX 15 12 Top-app capped at 12%
GPU_IDLE_TIMER 16 12 Faster GPU sleep
GPU_MAX_PCT 18 15 Minimal GPU ceiling
GPU_THERMAL_PWRLEVEL 6 7 One step lower than before
VM_SWAPPINESS 180 200 Maximum swap preference
VM_DIRTY_EXPIRE 180,000 240,000 Even lazier writeback
NET_TCP_RMEM/WMEM 4M 2M Smaller network buffers
WIFI_TXQLEN 64 32 Smallest viable TX queue
NET_TCP_KEEPIDLE 14,400 18,000 Longer keepalive = fewer radio wakes
NET_TCP_FIN 45 30 Faster connection cleanup

Expected result: lower idle drain during mixed day use, unchanged clean sleep behavior (V35 was already near-optimal there).

🌡️ Balanced Thermal Ceiling (New)

V35 logs showed a real issue: under a 9-hour heavy workload, balanced hit max_temp=81C with 251s in SUSTAINED. Balanced was still using the global sustained_temp_enter=65C, which is too hot for a "daily use" profile.

V36 adds per-profile sustained thermal thresholds for balanced:

Profile sustained_temp_enter sustained_temp_exit
Battery 65C (global) 52C (global)
Balanced 58C 50C
Performance 62C 54C

Expected: balanced enters SUSTAINED earlier under prolonged heavy load, keeping surface temperatures more comfortable during long sessions.

🔧 Install Fix

asb_install_prebuilt_governor was called on line 12 of common/install.sh — before the function was defined (line 109). Shell executes sequentially → not found error on every install.

Fix: Removed premature call. Added proper call after all function definitions, right before asb_big_banner.

🏚️ Removed Legacy Headroom<50 Shortcut

V35 introduced the unified sustained entry path:

throttle_signal -> throttle_confirmed -> warmup_grace -> debounce -> SUSTAINED

But a legacy block from V29 remained in asb_fsm.h:

if (headroom_pct > 0 && headroom_pct < 50) {
    desired = SUSTAINED;  // bypasses V35 logic entirely
}

This was a second secret entrance to SUSTAINED that skipped:

  • soft_clamp / hard_clamp classification
  • debounce (2-tick confirmation)
  • warmup grace (45s protection)
  • balanced thermal floor (48C)

Fix: Block removed. All SUSTAINED entries now go through the single V35 path.

⚙️ Fixed Global Load Thresholds

V35 had heavy_load_enter=2.0 and moderate_load_enter=10.0 — inverted. The safety guard silently clamped at runtime, but the config was logically wrong.

Threshold V35 (config) V35 (runtime) V36
moderate_load_enter 10.0 10.0 3.0
heavy_load_enter 2.0 10.5 (clamped) 5.0

Now performance has a proper ladder: MODERATE from load 3.0, HEAVY from load 5.0.

🔍 Honest Status JSON

In V35 heartbeat showed temp=45(stale) and sc=1 hc=0, but these fields were only in heartbeat (printed once every 15 minutes). A short diagnostic window never saw them.

V36 exposes the full thermal signal split in the status socket JSON:

{
  "state": "HEAVY", "profile": "performance",
  "temp": 45, "temp_valid": 0, "temp_age_s": 912,
  "soft_clamp": 1, "hard_clamp": 0,
  "thermal": 0, "headroom_pct": 67,
  ...
}

Four new fields: temp_valid, temp_age_s, soft_clamp, hard_clamp. Now every status query sees the complete picture.

🛠️ Reload Path History Bug

The socket reload command silently discarded running sessions when a profile change was detected — fsm_session_reset() was called without first writing history. Sessions that crossed a reload boundary were lost.

V36 fix: If a profile change is detected on reload, session_history_append_ex() is called before the reset, preserving the outgoing session with correct profile attribution.

📈 By The Numbers

Metric V35 V36 Change
Governor C lines 3,243 3,261 +18
FSM header 725 752 +27
Metrics header 456 472 +16
Config header 288 342 +54
Total C code 4,712 4,827 +115
Config parameters 50 57 +7
Status JSON fields 31 35 +4
Per-profile thermal tiers 2 3 +balanced
Performance CPU_CAP_BIG 3,456 MHz 3,264 MHz -200 MHz soft cap
Battery CPU_MAX_BIG 1,132 MHz 998 MHz -134 MHz hard cap
install.sh broken fixed Error gone

🏛️ Changed Files (from V35)

File What
common/install.sh Moved asb_install_prebuilt_governor call after definitions
profiles/performance.sh Rebuilt for sustainable 1-2h gaming (14 tuning changes)
profiles/battery.sh Deeper energy saving (16 tuning changes)
src/asb_fsm.h Removed legacy headroom<50 shortcut
src/asb_config.h Added balanced_sustained_temp_enter/exit
src/asb_governor.c Status JSON fields, reload history fix
src/asb_metrics.h Thermal signal split (V35 carried forward)
config/governor.conf moderate=3.0, heavy=5.0, balanced ceiling 58/50

Unchanged

system.prop, audio/camera/GPS overlays, balanced profile shell (already optimal per V35 logs), install scripts structure.

V35 fixed the thermometer and stopped balanced from treating every headroom dip as a five-alarm fire.
V36 rebuilt the gaming profile around 1-2h sustainable plateaus, pushed battery deeper into energy saving, gave balanced its own thermal ceiling, exposed honest telemetry everywhere, and fixed an install error that was hiding in plain sight since V27.

Assets 3
Loading