Hey Ralph,
Unfortunately I don't think this is going to work for us. Most of the
time we're starting the mpirun command using the ssh exec or shell
service, neither of which provide any mechanism for reading from file
descriptors other than 1 or 2. The only alternatives I see are:
1. Provide a separate command that starts mpirun at the end of a pipe
that is connected to the fd passed using the -xml-fd argument. This
command would need to be part of the OMPI distribution, because the
whole purpose of the XML was to provide an out-of-the-box experience
when using PTP with OMPI.
2. Implement an -xml-file option, but I could write the code for you.
3. Go back to limiting XML output to the map only.
None of these are particularly ideal. If you can think of anything
else, let me know.
Regards,
Greg
On Aug 30, 2009, at 10:36 AM, Ralph Castain wrote:
What if we instead offered a -xml-fd N option? I would rather not
create a file myself. However, since you are calling mpirun
yourself, this would allow you to create a pipe on your end, and
then pass us the write end of the pipe. We would then send all XML
output down that pipe.
Jeff and I chatted about this and felt this might represent the
cleanest solution. Sound okay?
On Aug 28, 2009, at 6:33 AM, Greg Watson wrote:
Ralph,
Would this be doable? If we could guarantee that the only output
that went to the file was XML then that would solve the problem.
Greg
On Aug 28, 2009, at 5:39 AM, Ashley Pittman wrote:
On Thu, 2009-08-27 at 23:46 -0400, Greg Watson wrote:
I didn't realize it would be such a problem. Unfortunately there is
simply no way to reliably parse this kind of output, because it is
impossible to know what the error messages are going to be, and
presumably they could include XML-like formatting as well. The
whole
point of the XML was to try and simplify the parsing of the mpirun
output, but it now looks like it's actually more difficult.
I thought this might be difficult when I saw you were attempting it.
Let me tell you about what Valgrind does because they have similar
problems. Initially they just had added --xml=yes option which
put most
of the valgrind (as distinct from application) output in xml
tags. This
works for simple cases and if you mix it with --log-
file=<filename> it
keeps the valgrind output separate from the application output.
Unfortunately there are lots of places throughout the code where
developers have inserted print statements (in the valgrind case
these
all go to the logfile) which means the xml is interspersed with
non-xml
output and hence impossibly to parse reliably.
What they have now done in the current release is to add a extra
--xml-file=<file> option as well as the --log-file=<file> option.
Now
in the simple case all output from a normal run goes well
formatted to
the xml file and the log file remains empty, any tool that wraps
around
valgrind can parse the xml which is guaranteed to be well
formatted and
it can detect the presence of other messages by looking for output
in
the standard log file. The onus is then on tool writers to look
at the
remaining cases and decide if they are common or important enough to
wrap in xml and propose a patch or removal of the non-formatted
message
entirely.
The above seems to work well, having a separate log file for xml
is a
huge step forward as it means whilst the xml isn't necessarily
complete
you can both parse it and are able to tell when it's missing
something.
Of course when looking at this level of tool integration it's
better to
use sockets that files (e.g. --xml-socket=localhost:1234 rather than
--xml-file=/tmp/app_XXXX.xml) but I'll leave that up to you.
I hope this gives you something to think over.
Ashley,
--
Ashley Pittman, Bath, UK.
Padb - A parallel job inspection tool for cluster computing
http://padb.pittman.org.uk
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
