[Darshan-users] Module contains incomplete data
Snyder, Shane
ssnyder at mcs.anl.gov
Wed Apr 27 10:20:59 CDT 2022
Thanks for working through the build issues and giving this a shot.
A couple of things stand out to me (ignoring your HDF5 issue for now):
* It looks like at least the MPI-IO module is no longer reporting partial data? Small progress...
* There is a new warning about there being no log utility handlers for a "null" module. Are you perhaps parsing a log generated by your prior Darshan install? Maybe you have not completely re-installed a new darshan-util? We should figure out what's going on there, too, to be safe.
I'd also suggest two things for your config file:
* Dial back MODMEM and MAX_RECORDS values. Your MODMEM value is asking Darshan to allocate a GiB of memory (it is expressed in MiB units and you set to 1024), which Darshan will happily try to do, I'm not sure it's a good idea though. I'd probably start with a MODMEM value of 8 and MAX_RECORDS of 2000, and just double those again if needed -- anything beyond that would be surprising unless you know your workload is really opening hundreds of thousands of files. You might also have a look at the files Darshan is currently instrumenting and see if you really want it to -- I've noticed when instrumenting Python frameworks that you can get tons of records for things like shared libraries, source files, etc. that can just be ignored using NAME_EXCLUDE mechanisms.
* Add "NAMEMEM 2" to your config file to force Darshan to allocate more memory (2 MiB) for storing the filenames associated with each record. This might actually be the main reason your log is reporting partial data rather than actually running out of module data, which is another reason not to get too aggressive with the MODMEM/MAX_RECORDS parameters. I should have mentioned this setting originally as there have been other users who have reported exceeding it recently.
Hopefully that gets you further along and we can move onto the HDF5 issue you mention.
Thanks,
--Shane
________________________________
From: Jiří Nádvorník <nadvornik.ji at gmail.com>
Sent: Wednesday, April 27, 2022 6:37 AM
To: Snyder, Shane <ssnyder at mcs.anl.gov>
Cc: darshan-users at lists.mcs.anl.gov <darshan-users at lists.mcs.anl.gov>
Subject: Re: [Darshan-users] Module contains incomplete data
Aha! I just realized there is an obvious "prepare.sh" script that I didn't run, I found out by trial and error though, could be more documented :).
Now I'm further.. With a config file:
MAX_RECORDS 102400 POSIX,MPI-IO,STDIO
MODMEM 1024
APP_EXCLUDE git,ls
I'm getting for:
darshan-parser --show-incomplete caucau_python_id127447-127447_4-27-48556-1842455298968263838_1.darshan |grep incomplete
output:
# *WARNING*: The POSIX module contains incomplete data!
# *WARNING*: The STDIO module contains incomplete data!
Warning: no log utility handlers defined for module (null), SKIPPING.
I don't think I have more than 100000 files to be touched by my poor tiny python script, right?
By the way I've encountered another problem, not sure whether to put it to another thread. If I compile with HDF5 (the results above are without it):
./configure --with-log-path=/gpfs/raid/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc --enable-hdf5-mod --with-hdf5=/gpfs/raid/SDSSCube/ext_lib//hdf5-1.12.0/hdf5/
It messes up my runtime and causes python to crash:
mpirun -x DARSHAN_CONFIG_PATH=/gpfs/raid/SDSSCube/darshan.conf -x LD_PRELOAD=/gpfs/raid/shared_libs/darshan/darshan-runtime/lib/.libs/libdarshan.so:/gpfs/raid/SDSSCube/ext_lib/hdf5-1.12.0/hdf5/lib/libhdf5.so -np 65 --hostfile hosts --map-by node /gpfs/raid/SDSSCube/venv_par/bin/python hisscube.py --truncate ../sdss_data/ results/SDSS_cube_c_par.h5
Resulting in:
INFO:rank[0]:Rank 0 pid: 137058
Darshan HDF5 module error: runtime library version (1.12) incompatible with Darshan module (1.10-).
Traceback (most recent call last):
File "hisscube.py", line 74, in <module>
writer.ingest(fits_image_path, fits_spectra_path, truncate_file=args.truncate)
File "/gpfs/raid/SDSSCube/hisscube/ParallelWriterMWMR.py", line 45, in ingest
self.process_metadata(image_path, image_pattern, spectra_path, spectra_pattern, truncate_file)
File "/gpfs/raid/SDSSCube/hisscube/CWriter.py", line 150, in process_metadata
h5_file = self.open_h5_file_serial(truncate_file)
File "/gpfs/raid/SDSSCube/hisscube/CWriter.py", line 170, in open_h5_file_serial
return h5py.File(self.h5_path, 'w', fs_strategy="page", fs_page_size=4096, libver="latest")
File "/gpfs/raid/SDSSCube/venv_par/lib/python3.8/site-packages/h5py-3.6.0-py3.8-linux-x86_64.egg/h5py/_hl/files.py", line 533, in __init__
fid = make_fid(name, mode, userblock_size, fapl, fcpl, swmr=swmr)
File "/gpfs/raid/SDSSCube/venv_par/lib/python3.8/site-packages/h5py-3.6.0-py3.8-linux-x86_64.egg/h5py/_hl/files.py", line 232, in make_fid
fid = h5f.create(name, h5f.ACC_TRUNC, fapl=fapl, fcpl=fcpl)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
File "h5py/h5f.pyx", line 126, in h5py.h5f.create
File "h5py/defs.pyx", line 693, in h5py.defs.H5Fcreate
RuntimeError: Unspecified error in H5Fcreate (return value <0)
You are saying that darshan should be compatible with HDF5 > 1.8, which 1.12 should be, right?
Thanks for help!
Cheers,
Jiri
st 27. 4. 2022 v 8:43 odesílatel Jiří Nádvorník <nadvornik.ji at gmail.com<mailto:nadvornik.ji at gmail.com>> napsal:
Hi,
I think I will chew through the documentation just fine but two things are not clear:
1. Does the darshan library provide its own config file that I need to change or do I need to always create my own?
2. How can I build the git version? I didn't find any instructions and the usual autoconf just throws:
* root at kub-b1:/gpfs/raid/shared_libs/darshan/darshan-runtime# autoconf
configure.ac:19<http://configure.ac:19>: error: possibly undefined macro: AC_CONFIG_MACRO_DIRS
If this token and others are legitimate, please use m4_pattern_allow.
See the Autoconf documentation.
configure.ac:21<http://configure.ac:21>: error: possibly undefined macro: AM_INIT_AUTOMAKE
configure.ac:22<http://configure.ac:22>: error: possibly undefined macro: AM_SILENT_RULES
configure.ac:23<http://configure.ac:23>: error: possibly undefined macro: AM_MAINTAINER_MODE
configure.ac:713<http://configure.ac:713>: error: possibly undefined macro: AM_CONDITIONAL
root at kub-b1:/gpfs/raid/shared_libs/darshan/darshan-runtime# ./configure
configure: error: cannot find install-sh, install.sh, or shtool in ../maint/scripts "."/../maint/scripts
Thanks for help.
Cheers,
Jiri
út 26. 4. 2022 v 17:43 odesílatel Snyder, Shane <ssnyder at mcs.anl.gov<mailto:ssnyder at mcs.anl.gov>> napsal:
Hi Jiri,
For some background, Darshan enforces some internal memory limits to avoid ballooning memory usage at runtime. Specifically, all of our instrumentation modules should pre-allocate file records for up to 1,024 files opened by the app -- if your app opens more than 1,024 files per-process, Darshan stops instrumenting and issues those warning messages when parsing the log file.
We have users hit this issue pretty frequently now, and we actually just wrapped up development of some new mechanisms to help out with this. They were just merged into our main branch, and we will be formally releasing a pre-release version of this code in the next week or so. For the time being, you should be able to use the 'main' branch of our repo (https://github.com/darshan-hpc/darshan) to leverage this new functionality.
There are 2 new mechanisms that can help out, both of which require you to provide a configuration file to Darshan at runtime:
* MAX_RECORDS setting can be used to bump up the number of pre-allocated records for different modules. In your case, you might try to bump up the default number of records for the POSIX, MPI-IO, and STDIO modules by setting something like this in your config file (this would allow you to instrument up to 4000 files per-process for each of these modules):
* MAX_RECORDS 4000 POSIX,MPI-IO,STDIO
* An alternative (or complementary) approach to bumping up the record limit is to limit instrumentation to particular files. You can use the NAME_EXCLUDE setting to avoid instrumenting specific directory paths, file extensions, etc by specifying regular expressions. E.g, the following settings would avoid instrumenting files with .so prefixes or files located in a directory we don't care about for all modules (* denotes all modules):
* NAME_EXCLUDE .so$ *
* NAME_EXCLUDE ^/path/to/avoid *
I'm attaching the updated runtime documentation for Darshan for your reference. Section 8 provides a ton of details on how to provide a config file to Darshan that should help clear up any missing gaps in my description above.
Please let us know if you have any further questions or issues, though!
Thanks,
--Shane
________________________________
From: Darshan-users <darshan-users-bounces at lists.mcs.anl.gov<mailto:darshan-users-bounces at lists.mcs.anl.gov>> on behalf of Jiří Nádvorník <nadvornik.ji at gmail.com<mailto:nadvornik.ji at gmail.com>>
Sent: Sunday, April 24, 2022 3:00 PM
To: darshan-users at lists.mcs.anl.gov<mailto:darshan-users at lists.mcs.anl.gov> <darshan-users at lists.mcs.anl.gov<mailto:darshan-users at lists.mcs.anl.gov>>
Subject: [Darshan-users] Module contains incomplete data
Hi All,
I just tried out Darshan and the potential output seems perfect for my HDF5 MPI application! Although I'm not able to get there :(.
I have a log that has a big stamp "This darshan log contains incomplete data".
When I run:
darshan-parser --show-incomplete mylog.darshan |grep incomplete
Output is:
# *WARNING*: The POSIX module contains incomplete data!
# *WARNING*: The MPI-IO module contains incomplete data!
# *WARNING*: The STDIO module contains incomplete data!
Would you be able to point me to some setting that would improve the measurements? Can I actually rely on the profiling results if it says the data is incomplete in some of the categories?
Thank you very much for your help!
Cheers,
Jiri
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/darshan-users/attachments/20220427/8a04a41d/attachment-0001.html>
More information about the Darshan-users
mailing list