Discussion of simplified IO500-script
by Julian Kunkel
Dear all,
based on the feedback of people and my own experience,
I have now drafted a simplified IO-500 run-suite:
https://github.com/VI4IO/io-500-script
I believe it is much easier to use, to debug and to optimize and quite minimal.
The design depends on phases:
1) A generation phase able to check for errors creates a hard-coded
run-script without branches. An example generated execution script is
enclosed.
=> during the generation phase you may use
https://www.vi4io.org/io500-info-creator/ to create the info tags.
It also creates a parallel minimal cleanup that purges the generated
files => again minimal.
2) The io-500.sh run-script can then be inspected / further modified
to include allowed tuning settings and the batch-submission commands
(which unfortunately were often missing in the last submissions).
3) An analysis phase which parses the output of the execution, checks
the validity of the result and computes the scores
Therewith, it minimizes the potential errors that may occur during the
*costly* execution and outsources them into the trivial preprocessing
and post-processing steps.
Outlook:
* I have not yet implemented step (3) or the build scripts, but will
do so depending on the feedback of the community.
* I reckon, the automatic detection of system information can be done
in the phases (1) as much as possible while some would need to go to
(2), but these calls can be added flexibly, too...
Best,
Julian
--
Dr. Julian Kunkel
Lecturer, Department of Computer Science
+44 (0) 118 378 8218
http://www.cs.reading.ac.uk/
https://hps.vi4io.org/
5 years, 3 months
New Pfind
by Julian Kunkel
Dear all,
after some testing (thx Ken) I just released a new Pfind
https://github.com/VI4IO/pfind
That one strips the libcircle dependency and adds some more features
for debugability.
Hopefully, that simplifies using IO-500 even further.
It has been tested with OpenMPI, MPICH and IntelMPI.
During testing it was a bit faster than the past version *and there is
now room for further improvement*.
Note that there IMHO is a way of gaming the value for find the way we
have setup IO-500, and I will give kudos to someone that actually
exploits it :-)
It is not trivial, though. The usage of the flags -C -P might be helpful.
You can test it without using IO-500 of course as well.
I e.g. run it with:
mpiexec -np 10 ./pfind /home/test/ -C -P -D rates
Best,
Julian
--
Dr. Julian Kunkel
Lecturer, Department of Computer Science
+44 (0) 118 378 8218
http://www.cs.reading.ac.uk/
https://hps.vi4io.org/
5 years, 4 months
IOR version 3.2.0 release candidate 1
by Glenn Lockwood
Dear I/O benchmarking community,
A release candidate for IOR version 3.2.0, the next major release of the
benchmark, is now available for download from the IOR GitHub repository:
https://github.com/hpc/ior/releases/tag/3.2.0rc1
It will be the first release of IOR to also include the mdtest benchmark
and includes a number of new features and back-end options. This release
branch has been in testing for over three months and should be quite
stable. Please download it, give it a test, and report any bugs or issues
you may encounter.
It is our goal to release 3.2.0 before the end of the year, so any feedback
you care to provide about this release candidate before then would be
greatly appreciated!
Sincerely,
The IOR maintainer team
5 years, 4 months
