[Trilinos-Users] [EXTERNAL] Belos::LinearProblem::setProblem() is taking too much time

[John Moore]

Thanks for the suggestion! It turned out to be a false alarm. I apologize.

I was measuring the time on the root process only, and it turns out I had a
function that was called before the call to Belos that is not easily
parallelizable. Therefore, some processes got to the Belos timing section
much faster than others, and I got a false reading on the root node. I
should probably use Teuchos::TimeMonitor from now on for timing things.



On Mon, Feb 16, 2015 at 2:42 PM, Thornquist, Heidi K <hkthorn at sandia.gov>
wrote:

>  Hi John,
>
>  The LinearProblem::setProblem method makes sure that the linear problem
> is valid, that there is an operator, right-hand side, and solution vector.
> After that, it computes the initial residual and preconditioned initial
> residual (if left preconditioning is used).  So, no, the behavior that you
> are seeing is not expected.
>
>  A couple questions might be helpful in understanding this better:
>
>  1)  Which solver are you using?
> 2)  How many iterations is this solver performing before achieving
> convergence (assuming 400 is your maximum number of iterations)?
> 3)  Are you deferring your custom preconditioner or operator construction
> until it is actually applied?  If you are, and you are using a left
> preconditioner, then it would show up in the setProblem timing.
>
>
>  Thanks,
> Heidi
>
>   --
>
>  Heidi K. Thornquist
>  Electrical Models & Simulation
> Sandia National Laboratories
> Albuquerque, NM  87185-1323
>
>
>   From: John Moore <johnpmooreiv at gmail.com>
> Date: Monday, February 16, 2015 12:10 PM
> To: "trilinos-users at software.sandia.gov" <
> trilinos-users at software.sandia.gov>
> Subject: [EXTERNAL] [Trilinos-Users] Belos::LinearProblem::setProblem()
> is taking too much time
>
>        I am interested in solving large linear systems within Newton
> iterations. My matrix structure remains constant, while only the values
> change.
>
>  I find that most of the time is actually spent in the functions
> Belos::LinearProblem::setProblem. Is this to be expected? The case that I
> am currently running is detailed below:
>
>  TpetraCrsMatrix
>  # processors:        32 mpi processes
>  Number of rows:   802225
>  Number of entries:  71819925
>  Preconditioner:       Custom
>  Number of iters:     400
>  Solve time:            42 Seconds
>  setProblem time:     35 seconds
>
>  The performance only gets work as the problem gets larger.  For a
> slightly larger (and denser) problem, the setProblem() time is about 4
> times that of the solve time. For small problems, most of the time is spent
> in the solve, as expected. Is there a way to get around calling
> setProblem() every time I need to solve the linear system, or could I be
> doing something wrong?
>
>  Thank you,
> John
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20150216/eb18e491/attachment.html>