Re: Login time

[Soeren Sandmann <sandmann daimi au dk>]

Federico Mena Quintero <federico ximian com> writes:

> El mié, 11-04-2007 a las 18:52 +0200, Soeren Sandmann escribió:
> > It will depend on the number of system calls per time of course, but
> > my experience is that strace times should not be trusted _at all_. In
> > one case I saw a factor of 6 difference between running with strace
> > and running without.
> 
> Hmm, interesting.
> 
> What I like about strace is that:

I'm not saying that strace is not useful, just that the timings it
reports are not accurate, and should not be mistaken for wall
time. 

What I think is going on is that strace greatly exaggerates the time
to do each system call so it tends to produce inflated numbers for
kernel time vs user time.

It might be interesting for someone to produce a kernel patch that
would log say the last 1000 system calls for a process and expose it
through /proc. That would produce much more accurate numbers than
ptrace can. (I'm assuming here that there is nothing dumb going on in
strace itself).


Soren