Opensees under Linux (Debian 9) slower than on Windows 7

[rafal]

I noticed a significant difference in execution time when running 'slopeAnalysis2Deff.tcl' (which requires velocityHistory.out) downloaded from [1] on the same computer but two operating systems. Namely under Windows the execution time is more than three times faster than on Linux Debian 9. On windows 7 (Dell Vostro, intel core i5 @ 2.50 GHz, RAM 6GB) the full analysis (5000 steps) finishes after 3.25 hours, but I also tested it with fewer steps, e.g. 450 steps in 390 seconds.
Any hints what causes such a difference?

[1] http://opensees.berkeley.edu/wiki/index ... s_of_Slope

[TPdct]

What version of OpenSees did you use on Windows 7? I mean, is it a compiled (by yourself) version or a downloaded version? Downloaded version of OpenSees is compiled by Intel Compiler and math computation is done by mkl (a math kernel library developed by Intel). That's why it's much faster.

[ja_abell]

On linux you need to turn on optimizations, native architecture (-march option) support and also provide a native BLAS or intel MKL for fastest support.

[rafal]

Thank you for hints. I tried various optimization flags when compiling recent Opensees (svn 6602) and the results are as follows:

Analysis execution time: 393 seconds - Windows 7, downloaded 2.5 version
Analysis execution time: 575 seconds - compiled with -Ofast
Analysis execution time: 964 second - compiled with -O3
Analysis execution time: 1376 seconds - compiled with -O2

Note that when I try to add -march=native (which is =sandybridge for Intel Core i5) the compilation process fails.

Any other specific ideas how to improve the performance of Opensees under Linux?

[fmk]

the release version of OpenSees is actually built with the intel compiler and linked to the mkl libraries for lapack and blas routines as opposed to the github versions of blas and lapack that are included in the default linux build .. if a student you can get the intel compilers for free from intel (these come with the mkl libraries) .. try building and linking with these,