cuDSS , MG mode and ILU(0)

Accelerated Computing GPU-Accelerated Libraries
,
1

Hello,

Recently, I tried cuDSS’s MG mode, and I was able to make it work for a sparse linear system (LU factorization, complex numbers) using up to 8 GPUs. I saw a reduction in factorization times and solution times. I also observed that the factors are distributed to the GPUs, which is also very good news.

My question would be this: I’m not sure if this is related to cuSPARSE or cuDSS. If the sparse matrix factorization and triangular matrix solution can be parallelized across GPUs (to a certain point), would it be possible to do it for ILU(0) decomposition, too? It is for the iterative solution of the system, and the sparse triangular matrix solution part is the most critical section, and it is not easy to implement straightforwardly. Any plans to implement preconditioner-type solvers too in the future?

(It could be useful when a linear system gets too large and it cannot fit into a single GPU’s memory)

Regards

Deniz

2

Hi Deniz, Thanks so much for the feedback. Have you looked into AmgX? This is a open source library of iterative solvers and pre-conditioners.

3

Hi Ariahi,

I have been aware of the AmgX package. I was able to compile it for Windows too, but I haven’t used it. I have been using cublas and cusparse libraries to build iterative solvers. I’m quite happy with them. I have never liked the algebraic multigrid approaches for some reason (geometric version too)

At the same time, I think I’ve found an answer to my question (a workaround). After ILU(0) decomposing a preconditioner matrix, those triangular matrices can be multiplied with each other, and this new matrix can be passed to cuDSS for factorization and solution using the MG mode. This could be a solution for now.

Cheers

Deniz