Skip to content

Latest commit

 

History

History
175 lines (127 loc) · 7.59 KB

README.md

File metadata and controls

175 lines (127 loc) · 7.59 KB

Optimising Micropython Boot time on ESP32 devices

Aim

To optimise the operation of micropython on ESP32 devices for operation as battery operated sensor devices. In particular, to investigate and optimise:

  1. Time (and power consumption) during boot from deepsleep

Summary of Results (see below for full data)

(Energy (uWh) = 1000 * Charge (mC) * 5V / (60 * 60))

Measuring time to wake FEATHERS3 (ESP32-S3) device from deepsleep, boot micropython and return to deepsleep:

Before Optimisation Time (ms) Charge (mC) Energy (microWh)
Boot to app_main() without Validation 47.0 1.7 2.4
Validation of Image 217.0 14.2 19.7
nvs_flash_init() 7.8 0.6 0.8
Allocate SPIRAM to GC 12.5 0.7 1.0
Micropython app_main() to _boot.py 0.8 0.1 0.8
_boot.py to main.py and deepsleep 350.0 23.6 32.8
Total: 635.1 41.9 58.3
After Optimisation Time (ms) Charge (mC) Energy (microWh)
Boot to app_main() without Validation 47.0 1.7 2.4
Micropython app_main() to _boot.py 0.8 0.06 0.1
_boot.py to deepsleep (_preboot.py) 1.0 0.06 0.1
Total: 48.8 1.82 2.5

For ESP32 device:

  • the optimised boot process takes 70ms (3.32mC) compared to 48.8ms (2.5mC) for the ESP32-S3

See ESP32 Wake Stubs for an additional power optimisation method.

Software

These experiments are based on my micropython experiments branch at:

Hardware

  • FeatherS3 device from UnexpectedMaker feathers3.io
    • Selected for low power consumption and support for my sensors.
  • Power Profiler Kit II to measure power consumption of device during experiments.
    • Also supports visualisation of timing pulses from device pins.
  • Cables, computer.

Measuring ESP32 boot times and power consumption

These measurements are made with the ESP32 device powered by USB (5V) and the PPK2 operating in ammeter mode to measure the current on the USB 5V line powering the ESP32 device. This method does not fully reflect the power efficiency of the device when running off a LiPoly battery, but is convenient for reprogramming and controlling the device during the test. The average current draw for the FeatherS3 device while in deepsleep is 1.73mA (at 5V USB).

See below for final power consumption measurements when the PPK2 is the power source (simulating a LiPo battery power source).

Full boot to main.py (unoptimised naive implementation)

main.py:

import time
import machine

pin = machine.Pin(18, machine.Pin.OUT)
pin.value(1)

# If we have just woken from a deep sleep go back to sleep
if machine.reset_cause() == machine.DEEPSLEEP_RESET:
    time.sleep_ms(10)
    pin.value(0)
    machine.deepsleep(1000)

Interpreting the images:

  • The greyed numbers show values for current, time and charge within the selected region.
  • The voltage levels for Pin 18 (the boot timing signal pin) are shown in the trace lablled 0 at the bottom of the image. These are used to calculate timing between key boot events in micropython.
Time to go back to sleep from main.py: 635ms. Power consumption: 41.3mC _
Time to start of micropython app_main(): 264ms (15.9mC) _
From app.main() to end of boardctl_startup() (nvs_flash_init()): 7.8ms (0.57mC) A
Time to execute gc_init() (with 16MB SPIRAM): 12.5ms (0.71mC) _
Time from call to load _boot.py to execution of deepsleep() in main.py: 350ms (23.6mC) A

Optimisation opportunities

A: Reduce time to boot to app_main(): (commit f2a3c66)

Add CONFIG_BOOTLOADER_SKIP_VALIDATE_IN_DEEP_SLEEP=y to ports/esp32/boards/sdkconfig.base:

  • Disables validation of the micropython image when the device wakes from deepsleep (see Espressif Docs))
Reduces time to boot to app_main() from 264ms to 47.0ms (15.9mC to 1.71mC) _
Total boot time back to sleep: 418ms (26.5mC) _

B: Execute app (deepsleep reboot) from frozen _boot.py module: (commit 8154c70)

Moving the code in main.py to ports/esp32/modules/_boot.py will eliminate the need to mount and initialise the filesystem on the flash memory before executing the app.
Time from _boot.py back to deepsleep reduced from 350ms (23.6mC) to 1.5ms (0.09mC) A
Total time from boot back to deepsleep: 70.4ms (3.17mC) A

C: Skip nvs_flash_init() on boot if waking from deepsleep: (commit 9bfa464)

NOTE: nvs_flash_init() must be called before initialising wifi (as RF calibaration data is read from nvs - EVEN if we disable NVS in the wifi config!!!).
Time for nvs_flash_init() reduced from 7.8ms (0.57mC) to 0ms (0mC). (200microseconds is time to generate pulse on logic pin). _

D: Skip initialisation of SPIRAM on boot if waking from deepsleep: (commit 450b5a3)

This reduces the work in gc_init() and is only of benefit on devices with SPIRAM.

NOTE: For most battery operated devices it is unlikely that additional SPIRAM would be required or be desirable, but I include it here as my testing device was equipped with SPIRAM.

NOTE: micropython no longer calls gc_init() on all available memory at boot (including PSRAM) and this change should no longer be necessary for v1.21+. (See micropython commit e465012 and this discussion).
gc_init() time reduced from 12.5ms (0.71mC) to less than 0.1ms. (200 microseconds is time to generate pulse on logic pin.) _

After all optimisations:
Time from boot to back to deepsleep: 50.0ms (1.87mC) (reduced from 264ms (15.9mC)). (Includes 6 * 0.2 = 1.2ms overhead for generating boot timing signals). _

Optimised code on ESP32 module

NOTE: This esp32 module board is not very power efficient (background current = 15.0mA), so take care in comparing with ESP32-S3 module above.
The ESP32 boot from, and return to, deepsleep takes 70ms (3.32mC) compared to 48.8ms (1.82mC) for the ESP32-S3. _