
Ublas : 
Subject: Re: [ublas] Question on ublas performance (resending as subscriber).
From: Nasos Iliopoulos (nasos_i_at_[hidden])
Date: 20100421 17:53:08
Thomas,
replace sparse_matrix with compressed_matrix. Also try the following:
use the generalized vector of compressed_vector A as in the link to build the stiffness matrix, I think this is faster than every other sparse container (mapped, compressed, coordinate). If you need a storage type like compressed (most likely if you use a classical sparse solver), just assign it after the generalized is filled by your stiffness matrix construction algorithm like:
compressed_matrix<double> K=A;
(I am unsure that in earlier versions of uBlas the above is optimum, the latest svn though should be ok. What version are you using?)
The downside of this approach is that it requires twice the memory.
Let us know if it works for you.
Best
Nasos
Date: Wed, 21 Apr 2010 13:51:56 0700
From: sgthomas27_at_[hidden]
To: ublas_at_[hidden]
Subject: Re: [ublas] Question on ublas performance (resending as subscriber).
Hi Jorn,
Your answer correctly identified the problem. It appears for my problem, mapped_matrix is a
much better (by factor of 110), if not the best choice as far as assembling goes. It also hasn't
hurt traversal that badly, it appears
In the link you attached (thank you!), the author uses something called "sparse_matrix" in
boost::numeric::ublas  does this even exist? Atleast in version 1.34, it gives me a compile
error, saying no sparse_matrix type exists in boost::numeric::ublas (it was the first thing I
tested before going for mapped_matrix) and the documentation online doesn't mention
sparse_matrix as a type of sparse storage at all.
Anyway, thanks a lot for all of your help!
Regards,
Sunil.
2010/4/21 Jörn Ungermann <j.ungermann_at_[hidden]>
Hi Sunil,
this is likely not a problem of uBLAS, but one of the principal problems
of using sparse matrices. Depending on the type of matrix either random
access or multiplication performance is efficient.
For the compressed_matrix, random access is rather costly *unless* you
can control the way in which elements are added to the matrix. If you
can assemble the (row_major) compressed_matrix rowbyrow with ascending
column indices, this should take no time at all.
If you can't do this, use a different matrix type for assembly, e.g.
mapped_matrix (which offers efficient random access, but bad
computational performance) and construct the compressed_matrix from
there.
See Gunter Winkler's page for details:
http://www.guwi17.de/ublas/matrix_sparse_usage.html
Kind reagrds,
Jörn
On Wed, 20100421 at 03:09 +0200, Sunil Thomas wrote:
> Hi all,
>
> I've been using boost 1.34 ublas library, especially the
> compressed_matrix class for sparse matrices in
>
> compressed row storage form. But I noticed that simply accessing an
> element of the matrix (to assign
>
> it a value, for example) slows my application down to unusable levels,
> for problems of the order of just
>
> 80,000 unknowns. I've identified the program is there and yes, I am
> allocating the memory as I should
>
> be for the matrix,  for example here is a snippet (of important
> lines):
>
>
> ************************************************************************************
> matrix_A = compressed_matrix(nelem_a(), nelem_a(), nonzeros()); //
> allocation
>
> matrix_A(uic1, uic1) += trans; // assignment
>
> matrix_A(uic2, uic2) += trans; // assignment
>
> ************************************************************************************
>
> where all variables (and/or functions), e.g. uic1, uic2, trans,
> neleme_a(), nonzeros(), etc.. are all welldefined
>
> (this is all been checked thoroughly). Commenting out the two
> assignment statements for example reduced
>
> my overall run time from 110 seconds to 0 (practically zero), for
> 80000 runs. Has anyone encountered this
>
> problem and know of a solution? I've heard a lot of stories about how
> boost::ublas is just not up there in
>
> performance and I certainly hope I am missing something trivial. Do
> later versions of boost address this
>
> better?
>
>
>
> Greatly appreciate any help.
>
> Thanks,
>
> Sunil.
>
>


Forschungszentrum Juelich GmbH
52425 Juelich
Sitz der Gesellschaft: Juelich
Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498
Vorsitzende des Aufsichtsrats: MinDir'in Baerbel BrummeBothe
Geschaeftsfuehrung: Prof. Dr. Achim Bachem (Vorsitzender),
Dr. Ulrich Krafft (stellv. Vorsitzender), Prof. Dr.Ing. Harald Bolt,
Prof. Dr. Sebastian M. Schmidt


_______________________________________________
ublas mailing list
ublas_at_[hidden]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: sgthomas27_at_[hidden]
_________________________________________________________________
The New Busy is not the too busy. Combine all your email accounts with Hotmail.
http://www.windowslive.com/campaign/thenewbusy?tile=multiaccount&ocid=PID28326::T:WLMTAGL:ON:WL:enUS:WM_HMP:042010_4