Lag Jacobian

[colinjcotter]

Hi folks,
I'm solving a linear time dependent PDE with a coefficient, and want to do lots of long runs with different values of the coefficient. To make this fast, I'm doing snes_lag_jacobian -2, so it reuses the LU factorisation. However, this means changing the coefficient doesn't do anything. Is there a way I can force it to update from the new coefficient between the runs?

all the best
--cjc

[wence-]

Hi folks, I'm solving a linear time dependent PDE with a coefficient, and want to do lots of long runs with different values of the coefficient. To make this fast, I'm doing snes_lag_jacobian -2, so it reuses the LU factorisation. However, this means changing the coefficient doesn't do anything. Is there a way I can force it to update from the new coefficient between the runs?

So you want it to be computed once in a solve (and reused for the entirety of that solve), but then the next solve you change the coefficient and so you want it recomputed? If so, then I think you want -snes_lag_jacobian_persists false

[colinjcotter]

No, I want it to be computed for a set of timesteps, and then reset it for a next set of timesteps.

…

________________________________________ From: Lawrence Mitchell ***@***.***> Sent: 08 June 2023 16:19 To: firedrakeproject/firedrake Cc: Cotter, Colin J; Author Subject: Re: [firedrakeproject/firedrake] Lag Jacobian (Discussion #2974) Hi folks, I'm solving a linear time dependent PDE with a coefficient, and want to do lots of long runs with different values of the coefficient. To make this fast, I'm doing snes_lag_jacobian -2, so it reuses the LU factorisation. However, this means changing the coefficient doesn't do anything. Is there a way I can force it to update from the new coefficient between the runs? So you want it to be computed once in a solve (and reused for the entirety of that solve), but then the next solve you change the coefficient and so you want it recomputed? If so, then I think you want -snes_lag_jacobian_persists false<https://petsc.org/main/manualpages/SNES/SNESSetLagJacobianPersists/> — Reply to this email directly, view it on GitHub<#2974 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ABOSV4QRHRPPFRZSCHRT6ZDXKHURHANCNFSM6AAAAAAY7NN4M4>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[wence-]

Ah, then just call solver.snes.setLagJacobian(-2) whenever you want it recomputed the next time snes calls snescomputejacobian.

[jaemoshin]

I've checked the attributes of snes and setLagJacobian does not seem to be an attribute.

[wence-]

I guess it is not yet exposed in petsc4py, you can probably figure it out by looking in petsc/src/binding/petsc4py/src/petsc4py/PETSc/SNES.pyx