src/case_studies/DasharoBenchRack_study.md: Add case study

[mkopec]

Add a case study for the BenchRack platform by 3mdeb. The BenchRack platform utilizes coreboot with LinuxBoot as the payload, replacing the UEFI firmware in its entirety with open source components (except for AMD PSP blobs)

[pamolloy]

Very cool, thanks for sharing! Just read through and looks good to me. Are there any benchmarks that could be made public? For example, do you measure/monitor boot time?

[mkopec]

Sure! Firmware takes 57 seconds to jump to the payload, half of which is spent in the PSP firmware. Once in the LinuxBoot payload, it takes a couple more seconds for the boot menu to appear.

ubuntu@3mdeb:~$ sudo cbmem -t
36 entries total:

   0:1st timestamp                                     29,106,318 (0)
  11:start of bootblock                                29,106,553 (235)
  12:end of bootblock                                  29,109,232 (2,678)
  13:starting to load romstage                         29,109,232 (0)
  17:starting LZ4 decompress (ignore for x86)          29,174,661 (65,428)
  18:finished LZ4 decompress (ignore for x86)          29,174,726 (65)
  14:finished loading romstage                         29,174,730 (3)
   1:start of romstage                                 29,174,734 (3)
   4:end of romstage                                   29,174,885 (151)
   8:starting to load ramstage                         29,174,885 (0)
  15:starting LZMA decompress (ignore for x86)         29,536,943 (362,058)
  16:finished LZMA decompress (ignore for x86)         29,583,677 (46,733)
   9:finished loading ramstage                         29,586,162 (2,485)
  10:start of ramstage                                 29,587,320 (1,158)
  30:device enumeration                                29,587,383 (62)
 112:started reading uCode                             29,587,766 (383)
 113:finished reading uCode                            29,617,222 (29,455)
  31:<unknown>                                         31,365,077 (1,747,855)
  40:device configuration                              31,371,820 (6,742)
  50:device enable                                     31,379,595 (7,775)
  60:device initialization                             31,381,611 (2,015)
  70:device setup done                                 31,884,811 (503,199)
 920:starting APOB read                                31,884,813 (2)
 921:starting APOB erase                               31,954,244 (69,431)
 922:starting APOB write                               39,357,767 (7,403,523)
 923:finished APOB                                     40,321,934 (964,166)
  75:cbmem post                                        40,343,709 (21,775)
  80:write tables                                      40,343,713 (3)
  85:finalize chips                                    40,352,487 (8,774)
  90:starting to load payload                          40,352,494 (6)
  15:starting LZMA decompress (ignore for x86)         54,335,177 (13,982,683)
  16:finished LZMA decompress (ignore for x86)         54,335,213 (35)
  15:starting LZMA decompress (ignore for x86)         56,242,188 (1,906,975)
  16:finished LZMA decompress (ignore for x86)         56,242,324 (136)
  15:starting LZMA decompress (ignore for x86)         56,242,327 (2)
  16:finished LZMA decompress (ignore for x86)         56,242,350 (23)
  99:selfboot jump                                     57,809,271 (1,566,921)

Stock ASRock UEFI can be measured with systemd-analyze, though that's likely excluding PSP time, which should be the same as in coreboot

ubuntu@3mdeb:~$ sudo systemd-analyze
[sudo] password for ubuntu:
Startup finished in 2min 58.748s (firmware) + 1.126s (loader) + 539ms (kernel) + 10.353s (initrd) + 10.407s (userspace) = 3min 21.174s
graphical.target reached after 10.402s in userspace.

So we go from 3:27 to 57 seconds :)

[pamolloy]

So we go from 3:27 to 57 seconds :)

Awesome! Would you mind adding it to the case study? This and any similar statistics are really valuable to show off the value of using LinuxBoot. Mention both the timing improvement and ease of use (i.e. using systemd-analyze rather than some custom script).

[mkopec]

Added benchmark section in 068acfe .

It should be noted that coreboot + LinuxBoot boot time cannot be measured using systemd-analyze, as it uses an EFI protocol to obtain firmware timestamps. But cbmem -t from the coreboot utils (available in many distros) is also easy.