<t>QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: mblattQuoteAnother issue is that matrices in FDM are of the order [200,500]^2 so I suspect LAPACK and ATLAS machinery is too heavy for this purpose. What's performance impact of hybrid C-LAPACK wrt a native solution in C++ or even Fo...

<t>QuoteAnother issue is that matrices in FDM are of the order [200,500]^2 so I suspect LAPACK and ATLAS machinery is too heavy for this purpose. What's performance impact of hybrid C-LAPACK wrt a native solution in C++ or even Fortran? There is function call overhead and where's the memory?Are you ...

You might want to advertise onmath-jobsphds.org.In math/scientific computing there are a lot of mailinglists with international audience, surely these exist in physics, too.

<t>Actually, I do not like discussions like "what is better: C++ or Fortran". I had to do some programming in FORTRAN a long time ago for a computational fluid dynamics course. After getting the basics (lines are only allowed to have a certain length, the first few characters of a line have a specia...

He passed away on Thursday, seeNY Times for an obituary.

BTW is the ADE code from your article available for download? I could not find it on your web site.

<t>As we collaborate(d) with both INRIA in France and Statoil Research Centre in Norway, I can assure you that at least some of the people doing reservoir simulation use algebraic multigrid methods for these kind of problems.BTW if it is not in the mkl, it does not mean that it is not there in other...

<t>QuoteOriginally posted by: zeta Central to many precon schemes is incomplete LU which *isn't* easily parallelizable. And that is why often domain decomposition methods or parallel multilevel schemes are used.QuoteI have yet to see fully parallel precon schemes. Would you care to mention what you ...

BTW: Jack Dongarra maintains a nice list of freely available linear algebra software.

<t>QuoteOriginally posted by: DcoleHowever, it needs to be parallelized. I am having a bit of difficulty with this. Does anyone have any links to a good discussion of the PCG algorithm's parts, as well as what needs to be shared when doing it across multiple processors? I basically have my x0 define...

<r>Probably very few domain decomposition experts.Why do you need books about domain decomposition methods in finance? I do not see what should be special about their application to finance.If you end up with a pde you can solve it in parallel using domain decomposition methods.If you need to know h...

<t>You said you are using PCG (preconditioned conjugate gradient). Therefore the preconditioner might want to use the diagonal entries. Due to the storage format I would suspect that referencing the diagonal triggers a binary search. This fails if the row is not sorted in the right way.But you shoul...

QuoteI don't agree that C++ code is always faster than Java, esp. with techniques like parallel computing etc.I agree in parts with the first part, as one can write slow C++ code (especially if one is a Java programmer.But what the hell has parallel computing got to do with it?

<t>Be reminded that it is allowed to actually store zero values as non-zeros in a sparse matrix!For the Dirichlet nodes you can do either of the following:1. Eliminate Dirichlet values from the linear system. I.e. They are not a degree of freedom anymore, but you need to change the values of the rig...

Out of curiosity: Why do you need to extract a submatrix to send it to another processor when doing red-black Gauss-Seidel?