Re: [Kwant] AUR build scripts for kwant and tinyarray
My last message only went to Michael, sorry for that, you can find it below.
I have now built mumps and all its dependencies (including scalapack) on Arch, which works fine for me.
How can I actually test whether Kwant uses mumps? The only warning I get is about my scipy missing umfpack. Is that the appropriate runtime warning? If not a runtime warning about a missing mumps might be a good idea. Also, is mumps detected at build or runtime or has the build.conf be filled manually?
best, Jörg
On 10.09.2015 17:08, Jörg Behrmann wrote:
On Thu, Sep 10, 2015 at 04:41:02PM +0200, Michael Wimmer wrote:
As for the PGP keys, I suppose
gpg --keyserver pgp.mit.edu --recv-key C3F147F5980F3535
is what is necessary, right?
Yep, that is exactly it.
There is one small issue though: mumps is listed as an optional dependency, and it sounds a bit as if kwant would be sped up if mumps is installed *after* kwant - but that wouldn't work, mumps has to be there on compile-time. Also, the mumps autodetect would not work in arch, I believe - so even with mumps installed kwant would not use it right now. This could be fixed with a appropriately filled-in build.conf, though (see http://www.kwant-project.org/doc/1.0/pre/install#build-configuration ).
That is something that I still wanted to test, as I don't have mumps currently installed.
I cannot test this, as one of the mumps dependencies (scalapack) does not compile for me at the moment. Maybe the best would be for now to remove mumps from the dependencies altogether?
I'll have have a look whether I can build mumps and its dependencies and if mumps has to be a compile-time dependency to be detected than mumps needs to be moved from optdepends to depends.
best, Jörg
Jörg Behrmann wrote:
How can I actually test whether Kwant uses mumps?
Try to “import kwant.solvers.mumps”. If this fails, the Kwant MUMPS solver does not work.
The only warning I get is about my scipy missing umfpack. Is that the appropriate runtime warning?
No, this is an independent thing.
If not a runtime warning about a missing mumps might be a good idea.
You might be right. MUMPS is not necessary for Kwant to work, and the performance of umfpack used to be quite OK for many applications. Now that umfpack is no longer included in SciPy, without MUMPS Kwant is indeed very slow.
Also, is mumps detected at build or runtime or has the build.conf be filled manually?
There is a very simple auto detection that works at least on Debian/Ubuntu [1]. In cases where this fails build.conf must be filled manually. If you can extend the auto detection to work on Arch as well, we will be more than happy to take this over.
Christoph
[1] http://git.kwant-project.org/kwant/tree/setup.py?h=stable&id=v1.0.4#n301
Hi everyone,
The only warning I get is about my scipy missing umfpack. Is that the
appropriate runtime warning?
If not a runtime warning about a missing mumps might be a good idea.
You might be right. MUMPS is not necessary for Kwant to work, and the
performance of umfpack used to be quite OK for many applications. Now that umfpack is no longer included in SciPy, without MUMPS Kwant is indeed very slow.
The current master of Kwant has a much clearer warning, which goes exactly what you're discussing, check it out and tell if you find it satisfactory. I removed the mention of umfpack: it's gone from scipy due to software rot. Instead I just complain when there's no mumps.
Best, Anton
Also, is mumps detected at build or runtime or has the build.conf be
filled manually?
There is a very simple auto detection that works at least on
Debian/Ubuntu [1]. In cases where this fails build.conf must be filled manually. If you can extend the auto detection to work on Arch as well, we will be more than happy to take this over.
Christoph
[1]
http://git.kwant-project.org/kwant/tree/setup.py?h=stable&id=v1.0.4#n301
Hey,
so that took me a bit longer to finally do the upload, but kwant and tinyarray can now be found again on the AUR [1,2]. Since the URL did not change, even the links on the Kwant website [3] work again. I'd be grateful if you could change that paragraph to
Arch install scripts for Kwant are kindly provided by Jörg Behrmann (formerly by Max Schlemmer). To install, follow the Arch User Repository installation instructions. Note that for checking the validity of the package you need to add the key used for signing to your user's keyring via
gpg --keyserver pool.sks-keyservers.net --recv-key C3F147F5980F3535
The fingerprint of the key is 5229 9057 FAD7 9965 3C4F 088A C3F1 47F5 980F 3535.
To answer the rest:
On Fri, Sep 11, 2015 at 03:35:26PM +0200, Christoph Groth wrote:
Jörg Behrmann wrote: How can I actually test whether Kwant uses mumps?
Try to “import kwant.solvers.mumps”. If this fails, the Kwant MUMPS solver does not work.
Ah, that's good to know. So from this I now get an ImportError: /usr/lib/libscotch.so: undefined symbol: gzclose when building kwant with mumps on Arch. I guess my libscotch is borked for some reason. I'll find out later why that is.
At least I can use this import to test whether the mumps configuration works and can later add a test to the PKGBUILD.
If not a runtime warning about a missing mumps might be a good idea.
You might be right. MUMPS is not necessary for Kwant to work, and the performance of umfpack used to be quite OK for many applications. Now that umfpack is no longer included in SciPy, without MUMPS Kwant is indeed very slow.
I just build the git head and the current warning is good.
best, Jörg
[1] https://aur.archlinux.org/packages/python2-kwant/ [2] https://aur.archlinux.org/packages/python2-tinyarray/ [3] http://kwant-project.org/install#arch-linux
Hi Jörg,
Thank you for updating the Arch packaging. I’ve updated the installation instructions following your suggestion.
Christoph
Hey Christoph,
much appreciated. Could you maybe add the link [1] on "Arch install scripts" (the first three words of the paragraph)?
best, Jörg
[1] https://aur.archlinux.org/packages/python2-kwant/ [2] http://kwant-project.org/install
On Wed, Sep 16, 2015 at 05:20:53PM +0200, Christoph Groth wrote:
Hi Jörg,
Thank you for updating the Arch packaging. I’ve updated the installation instructions following your suggestion.
Christoph
participants (3)
-
Anton Akhmerov -
Christoph Groth -
Jörg Behrmann