Parallel iterative solver (CG or FGMRES)

[bryce155]

I have a Incomplete Cholesky preconditioner and run the CG using RCI communication and it behaved very poor. THere is only a very little improvement from sequential and parallel mode. It took 80 secs for parallel and 86 for sequential. I am using intel Xeon X5650 2.67. Is it normal for iterative solver? I used the latest MKL 11. It scaled almost linear with direct solver (Pardiso)

Thanks,

Bryce

[Alexander_K_Intel2]

Hi, CG is RCI interface that doesn't affect performance of whole algorithm. Does your implementation of multiplication on stiffness matrix and precondition parallel or not? With best regards, Alexander Kalinkin

[bryce155]

Hi Alex, Thanks for the prompt response. I used 2 calls of mkl_dcsrtrsv in the preconditioner solve RCI=3 and mkl_dcsrsymv for matrix multiplication. Does it mean that those function dont perform well in parallel? Best regards, Bryce

[Gennady_F_Intel]

Bryce, yes, that's may be the problem: level 2 Sparse Triangular solvers (mkl_dcsrtrsv) is not threaded, but computing of m-v product of a sparse symmetrical matrix (mkl_dcsrsymv () ) is threaded. --Gennady

[bryce155]

Ok. Just one more question. Do you plan to include any parallel preconditioner for the iterative solver such as Block Jacobi (or block incomplete cholesky) or multigrid,etc?

[Gennady_F_Intel]

the only one thing I can say, that there are no such plans in the nearest release of MKL.

[bryce155]

Thank you Fedorov. I am still wondering the backward and forward substitutions in Pardiso were already parallel. Can we expect the same thing for the triangular solver in the near future? Thanks,

[Gennady_F_Intel]

Hello, there are no such plans in the nearest future. Gennady

[yanpu_z_]

I also noticed that when linking the parallel MKL libraries, the backward and forward substitutions (Ax=L*U*x=b<==>L*y=b, U*x=y) is almost the same as its sequential versions. Although the CPU usage is close to 100%, the speed for solving the equation A*x=b is not accelerated at all.

I also expect the triangular solver can be parallized in the near future.

Thanks very much!