Dear all, Is there a significant speed difference between MUMPS/scotch from Ubuntu repos vs. from source, especially considering that Kwant uses only the single core algorithms? Latest MUMPS in the Ubuntu repos is 4.10. Latest stable release of MUMPS is 5.0.5. I try to install everything from source. Speed is an issue for me. (Last work I did involved ~2M runs of a 4x4 hamiltonian defined on a 4x60000 wire with random potential per plot, which took ~16 days per plot with the computational power available to me.) However, it's easy (for me) to have a MUMPS installation go wrong, so I'd rather install them from repos unless there's a significant speed advantage. Thank you for your time and concern, Baris
Latest stable release of MUMPS is 5.0.5.
Sorry, that should have been 5.0.1.
Baris
On Tue, Sep 29, 2015 at 1:50 PM, Baris Pekerten
Dear all,
Is there a significant speed difference between MUMPS/scotch from Ubuntu repos vs. from source, especially considering that Kwant uses only the single core algorithms? Latest MUMPS in the Ubuntu repos is 4.10. Latest stable release of MUMPS is 5.0.5.
I try to install everything from source. Speed is an issue for me. (Last work I did involved ~2M runs of a 4x4 hamiltonian defined on a 4x60000 wire with random potential per plot, which took ~16 days per plot with the computational power available to me.) However, it's easy (for me) to have a MUMPS installation go wrong, so I'd rather install them from repos unless there's a significant speed advantage.
Thank you for your time and concern,
Baris
Thank you, Mike, will do. :)
On Tue, Sep 29, 2015 at 1:54 PM, Baris Pekerten
Latest stable release of MUMPS is 5.0.5.
Sorry, that should have been 5.0.1.
Baris
On Tue, Sep 29, 2015 at 1:50 PM, Baris Pekerten
wrote: Dear all,
Is there a significant speed difference between MUMPS/scotch from Ubuntu repos vs. from source, especially considering that Kwant uses only the single core algorithms? Latest MUMPS in the Ubuntu repos is 4.10. Latest stable release of MUMPS is 5.0.5.
I try to install everything from source. Speed is an issue for me. (Last work I did involved ~2M runs of a 4x4 hamiltonian defined on a 4x60000 wire with random potential per plot, which took ~16 days per plot with the computational power available to me.) However, it's easy (for me) to have a MUMPS installation go wrong, so I'd rather install them from repos unless there's a significant speed advantage.
Thank you for your time and concern,
Baris
Hi Baris,
From our experience, the MUMPS version does not make such a big difference though this may depend on the problem.
What can make a significant difference is which LAPACK/BLAS is used by kwant and MUMPS. Just from personal expericence I can tell that Intel's MKL is really good. Nowadays, as a researcher you can get a free license for MKL at https://software.intel.com/en-us/qualify-for-free-software/academicresearche... In order to use MKL, you need to install kwant from source, as discussed in the extended installation instructions. In particular, to use MKL you need to fill in a file called build.conf as explained here http://www.kwant-project.org/doc/1.0/pre/install#build-configuration You need not worry too much about you making mistakes in compiling it. You can always test the kwant installation by import kwant kwant.test() and I would guess that this should capture most errors that may arise from wrong linking, etc. Best, Michael On 29-09-15 12:50, Baris Pekerten wrote:
Dear all,
Is there a significant speed difference between MUMPS/scotch from Ubuntu repos vs. from source, especially considering that Kwant uses only the single core algorithms? Latest MUMPS in the Ubuntu repos is 4.10. Latest stable release of MUMPS is 5.0.5.
I try to install everything from source. Speed is an issue for me. (Last work I did involved ~2M runs of a 4x4 hamiltonian defined on a 4x60000 wire with random potential per plot, which took ~16 days per plot with the computational power available to me.) However, it's easy (for me) to have a MUMPS installation go wrong, so I'd rather install them from repos unless there's a significant speed advantage.
Thank you for your time and concern,
Baris
participants (2)
-
Baris Pekerten
-
Michael Wimmer