[petsc-users] [SLEPc] With the increasing of processors, result change in 1e-5 and solving time increase. krylovschur for MATSBAIJ matrix's smallest eigenvalue

[Jose E. Roman]

The computed eigenvalue has 7 matching digits, which agrees with the used tolerance. If you want more matching digits you have to reduce the tolerance.

The performance seems reasonable for up to 64 processes, so yes the problem may be too small for more processes. But performance depends also a lot on the sparsity pattern of the matrix. Take into account that scalability of matrix-vector products in SBAIJ matrices is expected to be worse than in AIJ. Anyway, to answer questions about performance it is better that you send the output of -log_view

Jose


> El 12 jul 2023, a las 10:18, Runfeng Jin <jsfaraway at gmail.com> escribió:
> 
> Hi,
>  When I try to increase the number of processors to solve the same matrix(to acquire the smallest eigenvalue) , I find all the results differ from each other within the  1e-5 scope (Though the ||Ax-kx||/||kx|| are all achieve 1e-8)  . And the solve time are first decreasing then increasing.
> my question is 
> (1) Is there anyway to make the result more constant 
> (2)  Does the time decreasing because the matrix dimension are too small for so many processors? Is there any way to reduce the solve time when increasing the number of processors?
> 
> Thank you!
> Runfeng
> 
> 
> matrix type  MATSBAIJ
> matrix dimension  2078802
> solver krylovschur
> blocksize 1
> dimension of the subspace   PETSC_DEFAULT
> number of eigenvalues   6
> the maximum dimension allowed for the projected problem   PETSC_DEFAULT
> -eps_non_hermitian 
> 
> number of processors		result	   solve time
> 16		-302.06881196 	526
> 32		-302.06881892 	224
> 64		-302.06881989 	139
> 128		-302.06881236 	122
> 256		-302.06881938 	285
> 512		-302.06881029 	510
> 640		-302.06882377 	291