• Jamboree18MemoBootchart

[original file is memo_ja-bootchart.txt]
[translated by ikoma]

CELF Technical Jamboree #18
Dec. 21, 2007
============================================================
CELF Technical Jamboree #18

06. Evaluation and Enhancement of Bootchart

============================================================

Evaluation and Enhancement of Bootchart

by Keisuke Yasui, Toshiba

* What is bootchart

• Generates graphs of processes and I/O load at boot
• Executes bootchartd script instead of /sbin/init

* Behavior analysis of bootchart

• Issues
  • Heavy
    • Takes 20 sec (5 sec in normal case)
  • Does not work with ash in busybox (no problem with bash)
  • Sparse sampling (200ms)
    • Higher sampling rate would make further heavier
  • Not easy to specify measurement interval
    • Up to invocation of getty
    • Up to receiving SIGUSER1
  • Not compatible with realtime priority process
    • 200ms sampling becomes almost 300ms; 1sec

* Enhancement

• Reimplemented bootcartd in C

  • 5 sec -> 6sec with 50ms sampling (load reduced considerably)

• Executes with highest realtime priority
• Starts in background at head of rcS
• Uses clock_gettime system call to acquire time
  • /proc/uptime was frequently called in original version
    • No influence even if time modifed with settimeofday()
• Better graph
  • Grids were too coarse with 5 sec interval; fixed to 1 sec
  • Coloring CPU usage
    • To use to tune so that heavy load work to be executed in part with light load
  • Only disks /dev/hd? and /dev/sd? were supported; analyzing software in Java modified to visualize others

* Remaining Issues

• To reduce load further
• Sampling intervals are not even
  • Deviates about 30ms as bootchart latency
    • This should be below 100ms[??]

* Q&A

• There exists bootchart.perl, which generates svg; no Java knowledge required
  • Java version works even on windows; convenient in some cases
• One of the headache is open/close of processes
  • Reimplementing in C was worthwhile