[Trilinos-Users] Memory usage when calling ComputePreconditioner()

[Eric Marttila]

Hello,

I'm solving a linear system multiple times, with the A matrix values changing 
each time.   Psuedocode is:

  //----------------------------------------------
  // create AztecOO solver and ML preconditioner 

  // Set AztecOO options
  solver.SetPrecOperator(MLPrec);
  solver.SetAztecOption(AZ_solver, AZ_cg);
  solver.SetAztecOption(AZ_output, 1);

  while (time < endTime) {
    // Modify the A matrix values.

    // Compute the preconditioner, now that A matrix valures are updated.
    MLPrec->ComputePreconditioner(true);

    // Solve the linear system.
    solver.Iterate(nIterations, 1e-10);
    time += timeStep;
  }
  //----------------------------------------------


I noticed a significant growth in memory usage upon each call to the 
ComputePreconditioner() function, and eventually my code would run out of 
memory.  I looked at the documentation of this ML function and saw that it 
recommends setting "reuse: enable" == true

I did this, and now the memory usage looks reasonable (i.e. it grows 
initially, then remains relatively constant throughout the remainder of my 
solve process).	

My question is this:  Is there any downside to setting "reuse: enable" to 
true?  (i.e. is this doing any sort of speed/memory tradeoff?)  For my test 
case I get about the same total solution time with and without this setting.  
One thing that has me wondering is that when I have "reuse: enable" set to 
true, I get additional output like this:

Time to check convergence rate = 7.11312 (s)

Thanks.
--Eric
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20130109/87521f27/attachment.html