[mvapich-discuss] Preloading the libraries "LD_PRELOAD" Effect

[Subramoni, Hari]

Hi,

The function of LD_PRELOAD is to make sure that functions defined in a shared library that is pointed to by LD_PRELOAD is selected before any similar definitions that are available/exposed by other shared libraries. It will not have an impact on startup cost.

If you have an application that uses several shared library objects and you want to improve the loading time of these, please look at the Spindle project from LLNL which is built exactly for this purpose (https://computing.llnl.gov/projects/spindle).

Depending on your application characteristics, there are several tuning/optimizations one can perform.

The following “Best Practices” link has some sample studies that we have done with applications highlighting the impact of application specific tuning on performance.

http://mvapich.cse.ohio-state.edu/best_practices/

For more details, please refer to the following presentation given at the “8th Annual MVAPICH User Group (MUG) Meeting” (http://mug.mvapich.cse.ohio-state.edu/mug/19/) about “How to Boost the Performance of Your HPC/AI Applications with MVAPICH2 Libraries?”

http://mug.mvapich.cse.ohio-state.edu/static/media/mug/presentations/19/subramoni-mug-19.pdf

Please let us know if you have any further questions and we will be glad to help.

Regards,
Hari.

From: mvapich-discuss-bounces at cse.ohio-state.edu <mvapich-discuss-bounces at mailman.cse.ohio-state.edu> On Behalf Of Kihang Youn
Sent: Wednesday, April 22, 2020 12:45 AM
To: mvapich-discuss at cse.ohio-state.edu <mvapich-discuss at mailman.cse.ohio-state.edu>
Subject: [mvapich-discuss] Preloading the libraries "LD_PRELOAD" Effect



Hi, everyone.

I'd like to know about the MVAPICH2 mpirun environment option "LD_PRELOAD".
Is the function of this option used to put the library on Cache (or RAM?) at startup and to gain benefits from multiple calls?

Use the LD_PRELOAD environment variable to export the path of the MVAPICH2 shared library object (libmpich.so or libmpi.so). eg: export LD_PRELOAD=/path/to/libmpi.so

Our team is looking for a way to load the library used by programs running through the "vmtouch" into the cache when performing the model, and we would like to know if it can be implemented through the "LD_PRELOAD" option.

The second question may be a lazy one.
Is there any options on tuning factor that can adjust parallel performance with MVAPICH2?
I'm actually only familiar with using MPI because I'm a newbie and I don't know much about performance enhancement, but I'm going to organize the list of recommendable performance-related options with MVAPICH2 and test them, so I'd like to know if there's anything organized(PDF or URL or something).

Thank you.



Kihang Youn(윤기항) - Application Analyst | Lenovo DCG Professional Services
Mobile: +82-10-9374-9396
E-mail: kyoun at lenovo.com<mailto:kyoun at lenovo.com>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20200427/d527e34a/attachment.html>