On Dec 15, 2013, at 11:59 PM, Ron Savage <[email protected]> wrote:

Does everyone know about this: http://research.swtch.com/pprof

Yes, but I’m skeptical about profiling libraries. At all.

1. Always prefer time-based sampling over call-based sampling.
call-based (like cachegrind or -pg) distorts the result heavily, and
there’s no need to do that.

2. You better let the kernel do the time-based sampling, not a library
(-lpprof, -lp) or instrumentation (cachegrind, dtrace)
 - which is basically run-time patching of every function call.
use dwarf to get the call-stack. afterwards, not online.

how they lie:

http://www.yosefk.com/blog/how-profilers-lie-the-cases-of-gprof-and-kcachegrind.html

sampling vs instrumenting

http://permalink.gmane.org/gmane.comp.lang.lua.luajit/3413


Just use the kernel provided framework: perf or systemtap on linux,
Instruments on darwin.

-- 
You received this message because you are subscribed to the Google Groups 
"marpa parser" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.

Reply via email to