[R] (performance) time in Windows vs Linux

[Zeljko Vrba]

On Tue, Jun 30, 2009 at 10:20:17AM +0900, Raymond Wan wrote:
> 
> I. Soumpasis wrote:
> >2009/6/29 C騷ar Freitas <cafanselmo12 at yahoo.com.br>
> >This is true. So I tried the same computer with windows XP and ubuntu 8.10
> >64bit dual core @3Gz and 4MB RAM
> >Windows 32bit results:
> >   user  system elapsed
> >  21.66    0.02   21.69
> >Linux 64bit Results
> >   user  system elapsed
> > 27.242   0.004  27.275
> >
> >This difference is small and it is truly explained by what the advanced
> >users have said.
> 
> 
> One minor comment which I forgot to mention is that a difference of 6 
> seconds for system that ran 30 seconds is worth noting, but may not be 
> statistically significant.  Especially when we're now talking about two 
> completely different OS' and, thus, two different ways of timing a 
> program.  If whatever data file you are using is also on a different 
> file system, then one could be fragmented, etc.
> 

Using wall-clock time metric is not "two different ways" of timing.  He could
have just as well measured the time using stop-watch, and the results would be
equally valid (given that his reaction time is ~0.5 seconds :-)) Also, "user"
time is time just spent in executing instructions in user-mode, which does
*not* account for waiting on disk I/O.  "system" time is time spent just in the
OS kernel, and that number would be also higher if there were intensive I/O
going on.  If I/O pauses were to blame (fragmentation, as you suggest, does
increase latency of serving disk requests), there would be a lot of idle time,
which would be manifested in a large difference between user+system and elapsed
(the latter being wall-clock time).

The numbers displayed above suggest that Windows version performs a *lot* less
work[*] than Linux version (6 seconds of CPU time is a  *lot* of work given
today's CPU frequencies).

[*] All programs actually run equally fast on a given machine, but some finish
    the same task sooner because they are doing less work...  

The answer can actually be obtained by compiling R with profiling enabled; this
will slow down execution considerably, but one will be able to pinpoint major
differences in performance.  Wild speculations about sources of this difference:
1) whatever compiler produced the Windows binary has better optimizations,
2) R heavily uses some standard library routines (within system-provided DLLs)
   that are better optimized under Windows than under Linux.