PaStiX vs. PARDISO performance

Dear CalculiX Users,

please find attached a graph illustrating the speedup with PaStiX for 9 representative examples (ranging from 700,000 to 5,000,000 degrees of freedom) involving face-to-face penalty nonlinear contact calculations. The graph is taken from the Master Thesis of Peter Wauligmann, who implemented the interface with PaStiX. The default implementation in CalculiX is “mixed”, which means first a direct solve in single precision followed by an iterative refinement in double precision. “GPU” means the installation of all packages described in the README-file, i.e. Cuda, Parsec, hwloc, scotch and PaStiX.

Best Greetings,

Guido

5 Likes

Is there any reference for this work which we can look into? Thanks.

This is a recent Master Thesis at the Technical University of Munich. I asked
Peter to upload his Thesis.

Best Greetings,

Guido

1 Like

Hello,
Could you clarify some points for understood research :

  • your CPU and GPU and amount of them.
  • vertical axe changed in limit from ~1.5 to 7.1, but not clear that is the 1. May be comparing PARDISO and PaStix with SPOOLES?
  • what the PARDISO https://www.pardiso-project.org/ or Intel® MKL PARDISO

Thank you.

1 Like

Hello konstantin,

  • Intel Xeon Gold 6244 (not 100% sure right now, but it’s an 8-core >4GHz), Nvidia V100 32GB
  • all you can see are speedups from PaStiX against PARDISO, so PARIDSO is 1
  • it is Intel MKL Pardiso

Dear CalculiX users,

I do not think Peter Wauligmann has uploaded his thesis yet, however, in the past he allowed me to send his thesis to whoever is interested. So just send a mail to dhondt@t-online.de and I will send you the pdf version.

Best Greetings,

Guido

1 Like

btw I did confirm, it was an Intel Xeon Gold 6244

I would be interested. Let me send an email. Thanks. Erhan