Hallöchen!
Does anybody have a howto for getting multicore support for SfePy? Something tested on Ubuntu would be ideal but is not necessary. I use umfpack, which needs a BLAS implementation. I suspect the BLAS determines whether all cores are used or only one.
umfpack's home page recommends GotoBLAS2. While Google finds some pages about installing GotoBLAS2 on Ubuntu, the problem is probably much large because umfpack needs to be told to use it.
By the way, a package called "libumfpack5.4.0" is installed on my system. Does SfePy use this? I remember that it installed its own umfpack code ...
Tschö, Torsten.
-- Torsten Bronger Jabber ID: torsten...@jabber.rwth-aachen.de or http://bronger-jmp.appspot.com
On 02/06/2012 10:24 AM, Torsten Bronger wrote:
Hallöchen!
Does anybody have a howto for getting multicore support for SfePy? Something tested on Ubuntu would be ideal but is not necessary. I use umfpack, which needs a BLAS implementation. I suspect the BLAS determines whether all cores are used or only one.
I am quite a new user of kubuntu, so cannot tell... It would be nice to be able to use multiple cores - maybe ask at the scipy mailing list? All we need is to make scipy/numpy use parallel blas.
umfpack's home page recommends GotoBLAS2. While Google finds some pages about installing GotoBLAS2 on Ubuntu, the problem is probably much large because umfpack needs to be told to use it.
Yes, I have no idea how to mix reasonably system-installed packages with your own. It might be better, if you feel like experimenting, to install everything locally. But it's certainly a lot of work.
By the way, a package called "libumfpack5.4.0" is installed on my system. Does SfePy use this? I remember that it installed its own umfpack code ...
If you use ubuntu 11.10 instructions from [1], umfpack will be used. The key package to install is libsuitesparse-dev. It works pretty fast (tests run faster than on gentoo on the same hw), but uses only a single core AFAIK.
r.
[1] http://docs.sfepy.org/doc-devel/installation.html#platform-specific-notes
Hallöchen!
Robert Cimrman writes:
On 02/06/2012 10:24 AM, Torsten Bronger wrote:
Does anybody have a howto for getting multicore support for SfePy? Something tested on Ubuntu would be ideal but is not necessary. I use umfpack, which needs a BLAS implementation. I suspect the BLAS determines whether all cores are used or only one.
I am quite a new user of kubuntu, so cannot tell... It would be nice to be able to use multiple cores - maybe ask at the scipy mailing list? All we need is to make scipy/numpy use parallel blas.
Before I ask on the scipy mailing list I must understand what you mean. So far, my SfePy uses its own version of umfpack, correct? So I must use another solver, namely a solver which uses some module in scipy/numpy. And then, my scipy/numpy installation must be made muticode-ready. ??
Tschö, Torsten.
-- Torsten Bronger Jabber ID: torsten...@jabber.rwth-aachen.de or http://bronger-jmp.appspot.com
On 02/06/2012 06:03 PM, Torsten Bronger wrote:
Hallöchen!
Robert Cimrman writes:
On 02/06/2012 10:24 AM, Torsten Bronger wrote:
Does anybody have a howto for getting multicore support for SfePy? Something tested on Ubuntu would be ideal but is not necessary. I use umfpack, which needs a BLAS implementation. I suspect the BLAS determines whether all cores are used or only one.
I am quite a new user of kubuntu, so cannot tell... It would be nice to be able to use multiple cores - maybe ask at the scipy mailing list? All we need is to make scipy/numpy use parallel blas.
Before I ask on the scipy mailing list I must understand what you mean. So far, my SfePy uses its own version of umfpack, correct? So I must use another solver, namely a solver which uses some module in scipy/numpy. And then, my scipy/numpy installation must be made muticode-ready. ??
No, sfepy uses umfpack via 'ls.scipy_direct' solver ('ls.umfpack' is its alias)
- it's just a wrapper setting up arguments and calling scipy.sparse.linalg.dsolve.spsolve().
Now scipy itself uses either the umfpack wrappers bundled with it (deprecated) or the umfpack scikit. In my kubuntu, the wrappers in scipy.sparse.linalg.dsolve.umfpack. Then the wrappers in turn use the /usr/lib/libumfpack.so from libsuitesparse-dev.
I hope it's clear(er) now :)
r.
participants (2)
-
Robert Cimrman -
Torsten Bronger