Dear useRs, I timed the same code (simulation with for loops) on the same box (dual Xeon EM64T, 1.5 Gb RAM) under 3 OSs and was surprised by the results: Windows XP Pro (32-bit): Time difference of 5.966667 mins 64-bit GNU/Linux (Fedora Core 4): Time difference of 6.966667 mins 32-bit GNU/Linux (FC4): Time difference of 9.2 mins (R 2.1.0 binaries downloaded from CRAN) I searched the archives and found out I should compile a non-SHLIB R binary. Is there something else I should pay attention to when trying to make R faster under GNU/Linux? Thank you, b.
On Mon, 2005-06-20 at 08:29 -0400, bogdan romocea wrote:> Dear useRs, > > I timed the same code (simulation with for loops) on the same box > (dual Xeon EM64T, 1.5 Gb RAM) under 3 OSs and was surprised by the > results: > Windows XP Pro (32-bit): Time difference of 5.966667 mins > 64-bit GNU/Linux (Fedora Core 4): Time difference of 6.966667 mins > 32-bit GNU/Linux (FC4): Time difference of 9.2 mins > (R 2.1.0 binaries downloaded from CRAN) > > I searched the archives and found out I should compile a non-SHLIB R > binary. Is there something else I should pay attention to when trying > to make R faster under GNU/Linux?On Fedora Core 4, try using gcc 3.4.4 (which you must install from source) instead of the default gcc 4.0.0. Martyn ----------------------------------------------------------------------- This message and its attachments are strictly confidential. ...{{dropped}}
On 20 June 2005 at 08:29, bogdan romocea wrote:
| I timed the same code (simulation with for loops) on the same box
| (dual Xeon EM64T, 1.5 Gb RAM) under 3 OSs and was surprised by the
| results:
| Windows XP Pro (32-bit): Time difference of 5.966667 mins
| 64-bit GNU/Linux (Fedora Core 4): Time difference of 6.966667 mins
| 32-bit GNU/Linux (FC4): Time difference of 9.2 mins
| (R 2.1.0 binaries downloaded from CRAN)
|
| I searched the archives and found out I should compile a non-SHLIB R
| binary. Is there something else I should pay attention to when trying
| to make R faster under GNU/Linux?
See 'help(Rprof)' for profiling support. It will tell how the time is
spent
by your loop, and I see no reason why you can't compare that across
architectures too.
Hth, Dirk
--
Statistics: The (futile) attempt to offer certainty about uncertainty.
-- Roger Koenker, 'Dictionary of Received Ideas of Statistics'