Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
6956 Discussions

Pardiso in OOC mode uses same memory as in IC mode

Pan11
New Contributor I
616 Views

I am trying to use pardiso in ooc mode with iparm(60)=2 from fortran.  It generates the ooc files but consumes the same ram as in ic mode iparm(60)=0.  Behavior is the same regardless of the size of the matrix, total or available mem, number or threads, or computer. Tested in centos 7 and 8 using 20.2 relase. Attached is the msglvl=1 output from runs for the same matrix.  Available mem scanned from /proc/meminfo is listed as well. 

Thanks for any help.

 

0 Kudos
1 Solution
Pan11
New Contributor I
566 Views

Hi Gennady, thanks for the info. 

 

Also, I have run several large tests and confirmed OOC mode with iparm(24)=1 consumes the memory estimated by iparm(63) mem in phase 22.  I will start another thread for the pardiso_handle_store issues.

 

Thank you!

View solution in original post

0 Kudos
5 Replies
Kirill_V_Intel
Employee
600 Views

Hello!

I don't see anything bad about the output you shared, which would explain what you see. One thing maybe: try to leave the swap size set to 0 rather than setting it to something else.

As a quick attempt, can you also try with iparm[23]=1 (so that you don't get the single-level factorization algorithm but rather a two-level one) and let us know the results?

Independent from whether this helps or not, I think we need a reproducer so that we can check what happens on our side.

 

Best,
Kirill

0 Kudos
Pan11
New Contributor I
597 Views

Hi Kirill, thank you for the quick response.

 

Swap size size does not have an effect.  However, setting iparm(24)=1 (I am using fortran) does reduce the memory consumption in OOC mode.  I will run larger matrices over the next few days and report behavior.

 

iparm(24)=1 causes segfault whith pardiso_handle_store, which I would also like to use.  Is there a way to use iparm(24)=1 with pardiso_handle_store?

 

Attached is a sample code and matrix data to reproduce the results.  I had to use a much smaller matrix than the original post due to the upload size limits.

 

 

0 Kudos
Gennady_F_Intel
Moderator
591 Views

regarding the uploading size limit -- we have an official support channel - Intel Online Service Center ( https://supporttickets.intel.com/servicecenter?lang=en-US) and You can open the ticket there against MKL. The uploading size limitation there <= 2Gb.


0 Kudos
Pan11
New Contributor I
567 Views

Hi Gennady, thanks for the info. 

 

Also, I have run several large tests and confirmed OOC mode with iparm(24)=1 consumes the memory estimated by iparm(63) mem in phase 22.  I will start another thread for the pardiso_handle_store issues.

 

Thank you!

0 Kudos
Gennady_F_Intel
Moderator
551 Views

The issue is closing and we will no longer respond to this thread.


0 Kudos
Reply