[petsc-users] Very slow SVD with SLEPC

[Jose E. Roman]

For Cross and TRLanczos, make sure that the matrix is stored in DENSE format, not in the default AIJ format. On the other hand, these solvers build the transpose matrix explicitly, which is bad for dense matrices in parallel. Try using SVDSetImplicitTranspose(), this will also save memory.

For SCALAPACK, it is better if the matrix is passed in the MATSCALAPACK format already, otherwise the solver must convert it internally. Still, the matrix of singular vectors must be converted after computation.

In any case, performance questions should include information from -log_view so that we have a better idea of what is going on.

Jose


> El 16 nov 2020, a las 6:04, Rakesh Halder <rhalder at umich.edu> escribió:
> 
> Hi Jose,
> 
> I'm only interested in part of the singular triplets, so those algorithms work for me. I tried using ScaLAPACK and it gives similar performance to Lanczos and Cross, so it's still very slow.... I'm still having memory issues with LAPACK and Elemental is giving me an error message indicating that the operation isn't supported for rectangular matrices. 
> 
> With regards to scaLAPACK or any other solver, I'm wondering if there's some settings to use with the SVD object to ensure optimal performance.
> 
> Thanks,
> 
> Rakesh
> 
> On Sun, Nov 15, 2020 at 2:59 PM Jose E. Roman <jroman at dsic.upv.es> wrote:
> Rakesh,
> 
> The solvers you mention are not intended for computing the full SVD, only part of the singular triplets. In the latest version (3.14) there are now solvers that wrap external packages for parallel dense computations: ScaLAPACK and Elemental.
> 
> Jose
> 
> 
> > El 15 nov 2020, a las 20:48, Matthew Knepley <knepley at gmail.com> escribió:
> > 
> > On Sun, Nov 15, 2020 at 2:18 PM Rakesh Halder <rhalder at umich.edu> wrote:
> > Hi all,
> > 
> > A program I'm writing involves calculating the SVD of a large, dense N by n matrix (N ~= 150,000, n ~=10,000). I've used the different SVD solvers available through SLEPc, including the cross product, lanczos, and method available through the LAPACK library. The cross product and lanczos methods take a very long time to compute the SVD (around 7-8 hours on one processor) while the solver using the LAPACK library runs out of memory. If I write this matrix to a file and solve the SVD using MATLAB or python (numPy) it takes around 10 minutes. I'm wondering if there's a much cheaper way to solve the SVD.
> > 
> > This seems suspicious, since I know numpy just calls LAPACK, and I am fairly sure that Matlab does as well. Do the machines that you
> > are running on have different amounts of RAM?
> > 
> >   Thanks,
> > 
> >      Matt
> >  
> > Thanks,
> > 
> > Rakesh
> > 
> > 
> > -- 
> > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> > -- Norbert Wiener
> > 
> > https://www.cse.buffalo.edu/~knepley/
>