Questions about intel package

LAMMPS LAMMPS Development
1

I have a couple of questions about intel package.

  1. Issue with lrt Keyword Interpretation:
    In the LAMMPS manual under the package command’s intel style, it’s stated that all of its settings, except the omp and mode keywords, are ignored if LAMMPS was not built with Xeon Phi co-processor support. However, in my recent test case, setting package intel 0 omp 3 lrt yes resulted in a runtime of 58 seconds, whereas using package intel 0 omp 3 lrt no gave a runtime of 1 minute 13 seconds. This discrepancy implies that the lrt keyword is having an effect. My setup involves running on an AMD EPYC 7452 32-Core Processor cluster, utilizing 64 cores per node. I employed 1 nodes (i.e. 64 cores), distributed among 16 MPI tasks, each using 4 cores. The LAMMPS version utilized is lammps-stable_2Aug2023. According to the manual, the lrt keyword should be ignored. Is there a potential misunderstanding in my interpretation?

  2. Compatibility Issue between run_style verlet/split and intel accelerator package:
    I encountered a Segmentation fault error when attempting to use run_style verlet/split along with the intel accelerator package. I was under the impression that these components should be compatible, yet the error persists.

I’ve attached my input file: test.in for the first question and test1.in for the second question. Both of them use sio2.data.

Your guidance would be greatly appreciated.

Jianghui Liu
test.in (428 Bytes)
test1.in (475 Bytes)
sio2.data (2.4 MB)

2

This is outdated and incorrect. As can be inferred from the description of the keywords, only the keywords “balance”, “ghost”, “tpc”, and “tptask” apply to using the INTEL package with Xeon Phi coprocessors.

The INTEL package is very special in that it does some additional initialization and copying of data to accommodate different precision settings and size and alignment requirements for vectorization. If I remember correctly, these steps would be triggered by the pair style. Yet, with run-style verlet/split the pair style is run in a different partition, so it would probably require some additional modifications to make it work with the intel package.

Can you please provide the exact command line that you are using in that case.

3

Thank you for your explanation.

The command line uesd in the second case is:

lmp_intel_cpu_intelmpi -p 48 16 -in test1.in

4

Thanks for the command line info. I have been able to reproduce the segmentation fault and also been able to determine its cause, and I can confirm that my suspicion is correct: some data structures internal to the INTEL packages are not initialized, since there was no INTEL package pair style run on the nodes where the PPPM class is run.

If you change your input to have:

suffix off
kspace_style    pppm 1e-5
kspace_modify   slab 3 diff ad
suffix on

It should run without crashing by using the non-intel version of pppm.