[R] sparse matrix tools

Martin Maechler maechler at stat.math.ethz.ch
Mon Jul 3 09:44:38 CEST 2006

>>>>> "JJTh" == Thaden, John J <ThadenJohnJ at uams.edu>
>>>>>     on Sun, 2 Jul 2006 07:38:21 -0500 writes:

    JJTh> Dear R-Help list: I'm using the Matrix library to
    JJTh> operate on 600 X ~5000 element unsymmetrical sparse
    JJTh> arrays. So far, so good, but if I find I need more
    JJTh> speed or functionality, 

Can you be more specific?  
   For asymmetric matrices, in our view, the most important gap to
   fill is the  LU  decomposition and hence solve() features.

Speed: Are you sure the time your R code spends is spent in 
       functions from "Matrix"?  {Did you use 'Rprof()' ?}
       If yes, which ones?

    JJTh> how hard would it be to
    JJTh> utilize other sparse matrix toolsets from within R,
    JJTh> say MUMPS, PARDISO or UMFPACK, that do not have
    JJTh> explicit R interfaces?  More information on these is
    JJTh> available here

    JJTh> www.cise.ufl.edu/research/sparse/umfpack/
    JJTh> www.computational.unibas.ch/cs/scicomp/software/pardiso
    JJTh> www.enseeiht.fr/lima/apo/MUMPS/ 

>From these, only the first one is open source.
Unfortunately, the PARDISO people seem to believe in money
making with scientific software -- a particular shame for since
they only work at most an hour away from me.
MUMPS is said to be "public domain", but then you only get it
after filling out a web form and only for a specific hardware.  Also,
it is about massively parallel computation, very interesting for
sparse matrices, but AFAIK not yet in our main focus.

UMFPACK is different, and even ships with the Matrix
package, because we have planned to interface to it, but haven't
yet got to that.  What parts of UMFPACK functionality would you
be interested in ?

    JJTh> and in these reviews

    JJTh> ftp://ftp.numerical.rl.ac.uk/pub/reports/ghsNAGIR20051r1.pdf
    JJTh> http://www.cise.ufl.edu/research/sparse/codes/ 

    JJTh> neither of which reviewed the R Matrix package, unfortunately.

    JJTh> Thanks, - John Thaden, Ph.D., U. Arkansas for
    JJTh> Med. Sci., Little Rock.

Martin Maechler,  ETH Zurich

More information about the R-help mailing list