Learning Forum

[MohammadMahdi]

Hi. I've recently moved from SMP-single precision to HPC-DMP-Intel MPI to reduce runtime. However, the results I 'm getting from DMP is highly different from SMP, though the k. files are exactly the same. Which of these two has to be trusted? The FEM/SPH model that I verified with SMP, and published, is now working differently using DMP and is giving me different results! I'm worried since I'm not able now to use my previously-verified model using DMP, though I've paid more money? How can I get the same trustable results as SMP using DMP?

Looking forward to hearing from you. Best.

[tslavik]

First, please assure you are using the same release of SMP and MPP so you have an apples-to-apples comparison. Some examples of release numbers are: R10.2.0, R11.2.2 and R12.1.0 - you will see this in the banner in the d3hsp file. And, please understand that R10.1.0 differs from R10.0.0, for example. We do see some differences between SMP and MPP, Their contact routines are coded differently, so that is typically the biggest cause of discrepancy.
To begin sorting out differences:
1/use double precision,
2/ use CONST=1 (see *CONTROL_PARALLEL) for SMP and *CONTROL_MPP_IO_LSTC_REDUCE for MPP.
3/ make sure the _MPP options of contact align with the SMP settings (e.g., IGNORE, BSORT, etc.)