Default
PaStiX
Pardiso
Spooles
Iterative scaling
Iterative Cholesky.
hi,
description below are extracted from document reference (p.538),
so, it depend on resource and how large is your model ? and right, in my experienced of large multipart contact analysis with plasticity. PaStiX is faster to convergence than MKL Pardiso.
I want to ask if all of these solvers are available in PPM because the last time I tried changing the default one (not sure if I attempted Spooles or Pardiso instead of Pastix) I received an error message saying that this solver was not found.
PaStiX and Spooles solver are already to run from official distribution. also for Pardiso, however it requires additional dynamic linking library (dll) from Intel MKL to be copied in the same directory.
libiomp5md
mkl_avx2.1
mkl_avx512.1
mkl_core.1
mkl_intel_thread.1
mkl_rt.1
mkl_mc3.1
Dear Synt san,
Thank you for your message!
so, it depend on resource and how large is your model ? and right, in my experienced of large multipart contact analysis with plasticity. PaStiX is faster to convergence than MKL Pardiso.
Could you let me know what MKL means ?
Dear lucas_bueno san,
Thank you for your message.
available in PPM
Could you let me know what PPM means ?
Dear synt san,
Thank you very much for your reply!!
Need PaStiX,Spooles,Pardiso use additional dynamic linking library (dll) as follows ?
Could you let me know how to get dll from intel MKL ?
Could you let me know where directory these dll be paste ?
libiomp5md
mkl_avx2.1
mkl_avx512.1
mkl_core.1
mkl_intel_thread.1
mkl_rt.1
mkl_mc3.1
no, it is only for Pardiso solver.
please follow the download link i posted before.
sub directories of installation folder
…\windows\redist\intel64_win\mkl
…\windows\redist\intel64_win\compiler
…\windows\mkl\lib\intel64_win
Dear synt san,
Thank you for your message !
I have already installed on my PC
Could you let me know which is dynamic linking library ?
Didn’t know it, thanks for the advice. I will install them.
It is PrePoMax abbreviature.
my notes above are for Intel MKL 2020/2021 versions. sorry i did not yet updates to the latest version (2022), but it seems the folder you shown only an extracted files not installation folder, is this right?
another easy method and straight forward are using ‘search’ feature of windows explorer in active or selected directories.
**edited
reading at another forums, for the latest version (2022) a substitute need change from 1 number to 2.
libiomp5md
mkl_core.2
mkl_def.2
mkl_intel_thread.2
mkl_rt.2 or mkl_rt.1(by renames)
mkl_mc3.2
mkl_avx2.2 or mkl_avx512.2 (chose one)
notes, environment variables setting plays an important role (e.g processor type Intel(gen) or AMD, large of RAM memory, drive type/speed and space)
@synt I tried the newer MKL library (2022.0.1). However, I get a segmentation fault error (see below). Have you tried the newer version at all?
[Thread 4600.0x1f18 exited with code 0]
Factoring the system of equations using the symmetric pardiso solver number of threads = 16
gdb: unknown target exception 0xc06d007e at 0x7ffd6ecc4fd9
Thread 1 received signal ?, Unknown signal.
0x00007ffd6ecc4fd9 in RaiseException () from C:\WINDOWS\System32\KernelBase.dll
i will try later, not frequently used Pardiso solver. being use PaStiX solver since early supported versions.
but looking of your error messages, it seems the library is incomplete to be copied (some missing).
Some information about the Pardiso integration can also be found on the @i_rokach page about PrePoMax: PrePoMax.
IMHO the best practical information about CCX solvers is available at this Mecway forum page. Here you can find detailed and relatively easy to follow instructions how to compile the Out-Of-Core Pardiso version.
@i_rokach i read the pages and graph by translate (Google), also took some comparison on my machine with Iterative Cholesky solver but running time not shown faster than PaStiX as describe.
another curious, did you manage to compared Out-Of-Core and Dynamically Linking versions of Pardiso solver?
It is typical for performance of iterative methods to be highly problem-dependent. In my case, a very simple static problem was solved. Due to this reason Cholesky solver was abnormally good. This is an educational web page and some examples are purely educational. When more real-life FE models are used, performance of CCX iterative solvers is usually poor. BTW, PaStiX solver is the mixed direct-iterative one. This is one of the reason it is so good for small and medium size problems.
I have never tried to compare OOC and in-core versions of Pardiso.
thanks for sharing experiences. i did not try personally at very big models, however another user with PaStiX subset “_i8” versions reported successfully for large problems.
Dear lucas_bueno san,
Thank you very much for your kindly reply!!!