Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
6652 Discussions

Pardiso cluster direct-iterative approach hangs

Rene_TT
Beginner
691 Views

I successfully use the direct-iterative approach on single machines for quite a while now, using the iparm[3] = 31 parameter. Now I am test-driving the cluster version and it hangs in the factorization phase when using this direct-iterative approach when running on more than one host. I observe this on Windows and Linux Centos7 machines, using oneAPI 2021.2.0. Before posting any details, my first question is: is this approach supposed to work for the cluster version? And if the answer is yes, is there an example somewhere illustrating this approach for the cluster version? That might already help me.

Thanks.

Labels (1)
0 Kudos
1 Solution
Kirill_V_Intel
Employee
664 Views

Hello!

1) Please refer to the iparm documentation for the cluster sparse solver: https://software.intel.com/content/www/us/en/develop/documentation/onemkl-developer-reference-c/top/...

You can see that iparm[3] is not used and there is no (except one typo I see) mentioning of direct-iterative approach for the cluster solver.

So, the feature is not supported for the cluster version.

2) I've been curious for a long time who is using this feature for PARDISO (non-cluster version). Could you describe a bit what is your use case, and how do you use this feature and why? I guess your matrix changes over time. Do you know if the changes are localized or contained within a certain subregion of the matrix or if the structure remains the same? 

The reason I'm asking is that we have several other features which may work better than direct-iterative approach.

 

Thanks,
Kirill

View solution in original post

5 Replies
MRajesh_intel
Moderator
675 Views

HI,


>>is there an example somewhere illustrating this approach for the cluster version?


You can refer to the MKL examples provided in /intel/oneapi/mkl/latest/examples. Here, you can extract the examples_cluster_c.zip and find cl_solver_unsym_c.c or other samples in c_mpi/cluster_sparse_solver/source directory.


Please let us know if you face any issues.


Regards

Rajesh.


Rene_TT
Beginner
671 Views

Hi Rajesh,

I used this example as well and it hangs to. I picked the example file 'cl_solver_unsym_c.c'  and add the following iparm setting to initiate the direct-iterative approach (multiple solving of structurally identical systems, first one with a direct solve, next ones iteratively):

iparm[ 3] = 31;

Then it hangs at the end of the first factorization phase when running with more than one mpi rank. Next the call on a Windows machine:

mpiexec -n 2 cl_solver_unsym_c.exe  

The same happens on the Linux machines (Centos7).

 

Kirill_V_Intel
Employee
665 Views

Hello!

1) Please refer to the iparm documentation for the cluster sparse solver: https://software.intel.com/content/www/us/en/develop/documentation/onemkl-developer-reference-c/top/...

You can see that iparm[3] is not used and there is no (except one typo I see) mentioning of direct-iterative approach for the cluster solver.

So, the feature is not supported for the cluster version.

2) I've been curious for a long time who is using this feature for PARDISO (non-cluster version). Could you describe a bit what is your use case, and how do you use this feature and why? I guess your matrix changes over time. Do you know if the changes are localized or contained within a certain subregion of the matrix or if the structure remains the same? 

The reason I'm asking is that we have several other features which may work better than direct-iterative approach.

 

Thanks,
Kirill

Rene_TT
Beginner
642 Views

Hi Kirill,

1) Thanks for the clear answer.  

2) Can I send you the details in a private conversation?

Regards,

Rene

Kirill_V_Intel
Employee
628 Views

Hi Rene!

A small update: we reproduced the hang on our side. While the feature is not supported, we will check if the hang can be fixed and get replaced by more meaningful behavior.

For your question: yes, it is absolutely fine to share details in the private conversation.

Best,
Kirill

Reply