Memory allocation issues with `cluster_sparse_solver` in MKL 11.3 with distributed matrix input

[Ivan_K_6]

Calling `cluster_sparse_sovler` with `-DMKL_ILP64`, where the input matrix is distributed with `iparm[39]=2`, produces the following:

*** Error in PARDISO  (     insufficient_memory) error_num= 1                                                                           
*** Error in PARDISO memory allocation: MATCHING_REORDERING_DATA, allocation of 1 bytes failed                                          
total memory wanted here: 75743 kbyte

 There is no effect from the number of nodes used for the MPI job.

Does the `cluster_sparse_solver` have to re-distribute the matrix between the nodes during the analysis phase?

I am not certain how to debug or diagnose this internal error. There is no issue with the amount of RAM on the machine.

The code compiles and runs under 11.2.3 as expected, the error only appears with MKL 11.3.

Thank you!

[Gennady_F_Intel]

Ivan, this is internal error and you couldn't diagnose this issue. How can we check problem on our side? Could you get the example to check on our side? Do you see the problem with LP64 version?

[Ivan_K_6]

Hello, Gennady.

The example code with compilation instructions is here: https://gist.github.com/anonymous/173e2dff21fef71bea51

I have also verified that the behavior with LP64 is identical.

Please let me know if there is anything else I could provide. Thanks!

[Gennady_F_Intel]

Ivan, I see you generate the artificial input and don't check the input matrix for correctness. Have you checked this?

[Ivan_K_6]

Setting `iparm[26] = 1` does not change the behavior.

[Gennady_F_Intel]

Yes, I agree. We see the problem on our side too. we are investigating the issue and will keep you informed when the fix of the problem would be released. Affected the 11.3 version of MKL.

[Gennady_F_Intel]

Ivan, the issue is fixed in 11.3. update 1. Please check how it works on your side.