Hardware/ARMEBS/3/LinuxFastBoot

Howto make Linux boot faster

This article is an howto which explain how to boot Linux faster on an ARMEBS3. More precisely, it talks about the boot time from -on to the time the kernel run it's init (= first userland task). After the first task, booting faster simply means running less task, or running them in parallel for optimizing CPU usage.

Current results

It's possible to have a shell after 2.59 seconds on the ARMEBS3, using kernel 2.6.19 (configured with support for all of our hardware), u-boot (r855) and a small romfs containing the (statically linked) busybox sh.

Timing details

• u-boot 0.328 seconds
• linux 1.656 seconds
• sh 0.606 seconds

Configuration

• Uncompressed kernel image (Image.img)
• u-boot variables:
  • stdout nulldev (be less verbose while booting)
  • verify n
    • Don't check the crc for the kernel image, flash is a reliable media, and can be checked once after it has been written, not on every boot.
  • bootargs "loglevel=4 init=/bin/sh lpj=350208 root=1f04 ro console=ttyS0,115200"
    • root file system on flash (1f04 == /dev/mtdblock4)
    • Be less verbose (loglevel)
    • Don't measure bogomips, use the value of lpj. If you don't change the clock and don't change the HZ value, the loop per jiffie won't change.

u-boot timing (as for r885)

• 3.8 Mo kernel image
• From reset to u-boot display_banner : 42 ms
  • was 90 ms
    • optimized PLL init in r875
    • optimized size of environment in r885
  • From banner to display_dram_config : 18 ms
    • was 200 ms
      • optimized size of environment in r885
  • From display_dram_config to "hit any key..." message : 3 ms
    • was 110 ms
    • reduce size of environment in r885
    • setenv does a *lot* of strcmp
    • replaced 10 ms active wait by wait on ready bit while accessing ethernet phy in r884
  • From "ok" to "Starting kernel" message : 235 ms
    • was 1320 ms
    • replaced memcpy, memmove, ... by linux assembly version (r874 and r880)

Measured speed as for r885

• memcopy from flash to ram
  • 235 ms for 3958672 bytes -> 16.84*10^6 bytes/seconds -> 16.06 Mo/s
• crc
  • 1.404 seconds for 3958672 bytes -> 2.82*10^6 bytes/seconds -> 2.68 Mo/s
• kernel uncompression (uncompressed size: 3958672)
  • speed: 1.556 seconds for 1956972 bytes -> 1.25*10^6 bytes/seconds -> 1.19 Mo/s