On Sun, Mar 4, 2012 at 9:14 AM, Roland Haas roland.haas@physics.gatech.edu wrote:
Hello Erik,
- /* if requested, change buffering mode of stdout */
You most likely have to (re-)set stderr to unbuffered after a possible freopen. freopen (at least from glibc, when I tested some time ago) sets the buffering to "full".
I didn't change the behaviour of stderr at all. Should we?
-erik
Hello Erik,
I didn't change the behaviour of stderr at all. Should we?
If I understand your patch correctly, then it makes sure that any buffering settings requested via the Cactus command line option are applied to all redirected stdout files (redirected via -redirect=oe) on the non-rank-0 MPI processes, right? However when stderr is redirected via freopen then its buffering settings are set to fully buffered while initially the C library opens stderr unbuffered. So while you did not change the behaviour with your patch, the patch seems to not quite do what is seems to intend to do, namely to make the C library buffering settings of stdout and stderr used for the CCTK_ProcXX.{out,err} files be the same as the ones on the rank-0 MPI process. So in that sense it would seem to make sense to change Cactus' current behaviour.
Yours, Roland
Implemented.
-erik
On Sun, Mar 4, 2012 at 5:42 PM, Haas, Roland roland.haas@physics.gatech.edu wrote:
Hello Erik,
I didn't change the behaviour of stderr at all. Should we?
If I understand your patch correctly, then it makes sure that any buffering settings requested via the Cactus command line option are applied to all redirected stdout files (redirected via -redirect=oe) on the non-rank-0 MPI processes, right? However when stderr is redirected via freopen then its buffering settings are set to fully buffered while initially the C library opens stderr unbuffered. So while you did not change the behaviour with your patch, the patch seems to not quite do what is seems to intend to do, namely to make the C library buffering settings of stdout and stderr used for the CCTK_ProcXX.{out,err} files be the same as the ones on the rank-0 MPI process. So in that sense it would seem to make sense to change Cactus' current behaviour.
Yours, Roland
users@lists.einsteintoolkit.org
-
Erik Schnetter -
Haas, Roland