You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This issue started with #7095. However, it has grown to encompass several things, so I'm opening a new issue to gather them all into one spot.
The initial issue is that mpirun's --output-filename behavior no longer matches what is described in the mpirun(1) man page. Specifically: it doesn't just output to a single file per process any more; mpirun now creates a directory for each MPI process and outputs a stdout and stderr file in there.
We should probably rename this behavior to be --output-directory.
The new behavior needs to be documented: output to DIR/JOBID/rank.N/stdout|stderr
Can also be combined with --merge-stderr-to-stdout (right? test this...)
Be sure to mention :nojobid (omit the job ID in the directory hierarchy) and :nocopy (don't also emit to stdout/stderr), and that they can be combined into a single comma-delimited list
Mention efficiency of :nocopy (i.e., no IOF used to send back to mpirun) -- probably nearly as efficient as the app writing to its own local files...?
BUG FIX: If you specify an invalid suffix (e.g., :noooooocopy, the user is not notified).
Need to think through how to use this feature over time from user's perspective
What to do in v2.x
What to do in v3.0.x (behavior changed compared to v2.x)
What to do in v3.1.x
What to do in v4.0.x (?added -output-directory / deprecated --output-filename?)
What to do in v5.0.x (?--output-directory only?)
Make sure to mention killing/deprecating --output-file in NEWS
...unless the old --output-file behavior is resurrected / preserved...? That's an option, if someone wants to do it.
The text was updated successfully, but these errors were encountered:
What does nocopy does?
If it only writes to files but not stdout/stderr then we still need to involve IOF since the MPI tasks might be running on nodes not mounting the filesystem mpirun is writing too.
For performance improvements, I think we would need an other option such as noforward and this should probably not be the default behavior.
This issue started with #7095. However, it has grown to encompass several things, so I'm opening a new issue to gather them all into one spot.
The initial issue is that
mpirun
's--output-filename
behavior no longer matches what is described in thempirun(1)
man page. Specifically: it doesn't just output to a single file per process any more;mpirun
now creates a directory for each MPI process and outputs astdout
andstderr
file in there.--output-directory
.DIR/JOBID/rank.N/stdout|stderr
--merge-stderr-to-stdout
(right? test this...):nojobid
(omit the job ID in the directory hierarchy) and:nocopy
(don't also emit to stdout/stderr), and that they can be combined into a single comma-delimited list:nocopy
(i.e., no IOF used to send back tompirun
) -- probably nearly as efficient as the app writing to its own local files...?:noooooocopy
, the user is not notified).-output-directory
/ deprecated--output-filename
?)--output-directory
only?)--output-file
in NEWS--output-file
behavior is resurrected / preserved...? That's an option, if someone wants to do it.The text was updated successfully, but these errors were encountered: