RI_MP2 memory

[Niamh O'Neill]

Dear CP2K developers,

I am attempting to compute energies and forces for a system containing 6 Na Cl ions each and 52 waters at RI_MP2 level.

I am having issues in terms of memory, and get a segmentation fault for invalid memory reference.

The system I am running on has 256 GB RAM per node and I am running on 32 nodes . (128 cores per node and have tried both 64 and 32 tasks per node) and I start from a PBE wavefunction.

Do you know why these issues might be happening?

I have attached my input files below.

Thank you in advance for the help.

Best wishes,

-Niamh

[Augustin Bussy]

Hi Niamh,

it is hard to know where things go wrong without your output file. Is it during the SCF or MP2? Generally, I see that you are using a RI_METRIC of type identity. This does not bring much for RI_MP2. The use of a RI metric helps to create sparsity, which is only really exploited in low-scaling algorithms.

You can try setting the keyword WF_CORRELATION%CONONICAL_GRADIENTS%FREE_HFX_BUFFER to FASLE. In the past, this was known to create crashes with MP2 gradients. Depending on the CP2K version you are using, this keyword might be in a different input section.

Best,

Augustin

---

From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of Niamh O'Neill <niamh9...@gmail.com>
Sent: Wednesday, June 14, 2023 7:57 PM
To: cp2k <cp...@googlegroups.com>
Subject: [CP2K:18928] RI_MP2 memory

 

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/cf2610e9-22be-4829-a355-f04729d1e7e0n%40googlegroups.com.

[Frederick Stein]

 Hi Niamh,

Apart from Augustin's hint, I have some more ideas:

Your machine should have more than enough memory. We run calculations with systems of similar size on Piz Daint (with much less memory than you). Please provide the output file, without it, we can just guess.

Can you maybe reduce the number of ranks per node?

If it breaks in the integral calculation part, you could decrease the available memory for MP2 (MEMORY keyword in the WF_CORRELATION section). You will not need that much memory.

But I agree with Augustin, without any further information, it is difficult to narrow it down to the actual issue as you apparently have more than enough memory available.

Best,

Frederick

[Niamh O'Neill]

Thank you both Augustin and Frederick for your replies.

The issue had been after the MP2 energy had been computed and so on the gradient step.

Setting WF_CORRELATION%CONONICAL_GRADIENTS%FREE_HFX_BUFFER to FALSE seems to have done the trick and it is now all working as expected.

Thank you very much for the help.

Best wishes,

-Niamh 

You received this message because you are subscribed to a topic in the Google Groups "cp2k" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/cp2k/mJtHrSg9Kuw/unsubscribe.
To unsubscribe from this group and all its topics, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/GV0P278MB0370DFA61B6B344EC37E02DFD35BA%40GV0P278MB0370.CHEP278.PROD.OUTLOOK.COM.