Have experience with linear algebra libraries for large sparse SPD systems?

[eLiu]

Hi all,
I need to solve a large linear system, often in parallel since the number of unknowns won't fit on one computer. The code should also work in serial on smaller systems. So if the I ask for cholesky on 1 processor, the performance should be like in LAPACK... if not, let me know and I'll call out to LAPACK instead.

**The matrix is SPD.** (It's from a linear elasticity problem discretized with CG finite elements.)
**I'm coding in C.** (The rest of the project is in C; this is not negotiable.)
**This is a temporary solution, so it needs to work correctly & not be horribly slow. But I don't have long to set it up, so ease of implementation is a plus.**

I would like a linear algebra library/package (or set of of libraries/packages) that can cover the following capabilities:
1) reasonable speed in parallel AND serial*
2) iterative solver (Conjugate Gradients) AND preconditioners (at least ILU(0))
3) Cholesky with the option of reordering for sparsity (e.g. nested dissection)
4) Easy to interface with C (i.e. written in like F77 or C).

Could you guys suggest/share experiences with libraries/packages that provide my desired features? I'd prefer only dealing with 1 'jack-of-all-trades' package, but if picking up a few specialized ones is better, that's fine too.

I've been searching around, and it seems like the most comprehensive package is going to be PETSc. I have 0 experience with PETSc but people in my group don't like it, claiming that it carries around a lot of overhead. Anyone have experience working with PETSc (in terms of performance & ease of use)?

Thanks,
-Eric

 

[slugg]

I have some limited experience with PETSc from a Wavelet Transformations class I took a year ago. I used the matrix structure/interface from PETSc to implement some transformations for a plugin to Mathematica. It's been a while, but I remember the documentation to be very easy to follow, the interfaces are clean and consistent, and it comes with all sorts of examples. It's a huge library, so it's kind of overwhelming.

Personally, I didn't like it. Immediately after that project, I switched to J-Link to integrate Mathematica with Java, then did all my implementation in Java. Performance took a huge hit, but I could actually finish my work on time...

I haven't gotten to your level of computational mathematics, so I can't make a direct recommendation for you. I've only been here for 2 years...

Good luck!