[petsc-users] Parallel TS for ODE
Stefano Zampini
stefano.zampini at gmail.com
Tue Apr 20 15:07:18 CDT 2021
>
>
> Thank you for the advices, I would just like to convert the code I already have to see what might happen once parallelized.
You are not really listening to our advices. I can tell you what happens to 3 coupled ODEs split on 3 processes. The solver will be slower, by far.
> Do you think it is better to put the 3 equations into a 1d Distributed Array with 3 dofs and run the job with multiple procs regardless of how many equations I have? Is it possible?
>
> Thank you,
> Francesco
>
>> Il giorno 20 apr 2021, alle ore 17:57, Stefano Zampini <stefano.zampini at gmail.com <mailto:stefano.zampini at gmail.com>> ha scritto:
>>
>> It does not make sense to parallelize to 1 equation per process, unless that single equation per process is super super super costly.
>> Is this work you are doing used to understand PETSc parallelization strategy? if so, there are multiple examples in the sourcetree that you can look at to populate matrices and vectors in parallel
>>
>> Il giorno mar 20 apr 2021 alle ore 17:52 Francesco Brarda <brardafrancesco at gmail.com <mailto:brardafrancesco at gmail.com>> ha scritto:
>> In principle the entire code was for 1 proc only. The functions were built with VecGetArray(). While adapting the code for multiple procs I thought using VecGetOwnershipRange was a possible way to allocate the equations in the vector using multiple procs. What do you think, please?
>>
>> Thank you,
>> Francesco
>>
>>> Il giorno 20 apr 2021, alle ore 16:43, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> ha scritto:
>>>
>>> On Tue, Apr 20, 2021 at 10:41 AM Francesco Brarda <brardafrancesco at gmail.com <mailto:brardafrancesco at gmail.com>> wrote:
>>> I was trying to follow Barry's advice some time ago, but I guess that's not the way he meant it. How should I refer to the values contained in x? With Distributed Arrays?
>>>
>>> That is how you get values from x. However, I cannot understand at all what you are doing with "mybase".
>>>
>>> Matt
>>>
>>> Thanks
>>> Francesco
>>>
>>>>> Even though it will not scale and will deliver slower performance it is completely possible for you to solve the 3 variable problem using 3 MPI ranks. Or 10 mpi ranks. You would just create vectors/matrices with 1 degree of freedom for the first three ranks and no degrees of freedom for the later ranks. During your function evaluation (and Jacobian evaluation) for TS you will need to set up the appropriate communication to get the values you need on each rank to evaluate the parts of the function evaluation needed by that rank. This is true for parallelizing any computation.
>>>>>
>>>>> Barry
>>>
>>>
>>>
>>>> Il giorno 20 apr 2021, alle ore 15:40, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> ha scritto:
>>>>
>>>> On Tue, Apr 20, 2021 at 9:36 AM Francesco Brarda <brardafrancesco at gmail.com <mailto:brardafrancesco at gmail.com>> wrote:
>>>> Hi!
>>>> I tried to implement the SIR model taking into account the fact that I will only use 3 MPI ranks at this moment.
>>>> I built vectors and matrices following the examples already available. In particular, I defined the functions required similarly (RHSFunction, IFunction, IJacobian), as follows:
>>>>
>>>> I don't think this makes sense. You use "mybase" to distinguish between 3 procs, which would indicate that each procs has only
>>>> 1 degree of freedom. However, you use x[1] on each proc, indicating it has at least 2 dofs.
>>>>
>>>> Thanks,
>>>>
>>>> Matt
>>>>
>>>> static PetscErrorCode RHSFunction(TS ts,PetscReal t,Vec X,Vec F,void *ctx)
>>>> {
>>>> PetscErrorCode ierr;
>>>> AppCtx *appctx = (AppCtx*) ctx;
>>>> PetscScalar f;//, *x_localptr;
>>>> const PetscScalar *x;
>>>> PetscInt mybase;
>>>>
>>>> PetscFunctionBeginUser;
>>>> ierr = VecGetOwnershipRange(X,&mybase,NULL);CHKERRQ(ierr);
>>>> ierr = VecGetArrayRead(X,&x);CHKERRQ(ierr);
>>>> if (mybase == 0) {
>>>> f = (PetscScalar) (-appctx->p1*x[0]*x[1]/appctx->N);
>>>> ierr = VecSetValues(F,1,&mybase,&f,INSERT_VALUES);
>>>> }
>>>> if (mybase == 1) {
>>>> f = (PetscScalar) (appctx->p1*x[0]*x[1]/appctx->N-appctx->p2*x[1]);
>>>> ierr = VecSetValues(F,1,&mybase,&f,INSERT_VALUES);
>>>> }
>>>> if (mybase == 2) {
>>>> f = (PetscScalar) (appctx->p2*x[1]);
>>>> ierr = VecSetValues(F,1,&mybase,&f,INSERT_VALUES);
>>>> }
>>>> ierr = VecRestoreArrayRead(X,&x);CHKERRQ(ierr);
>>>> ierr = VecAssemblyBegin(F);CHKERRQ(ierr);
>>>> ierr = VecAssemblyEnd(F);CHKERRQ(ierr);
>>>> PetscFunctionReturn(0);
>>>> }
>>>>
>>>>
>>>> Whilst for the Jacobian I did:
>>>>
>>>>
>>>> static PetscErrorCode IJacobian(TS ts,PetscReal t,Vec X,Vec Xdot,PetscReal a,Mat A,Mat B,void *ctx)
>>>> {
>>>> PetscErrorCode ierr;
>>>> AppCtx *appctx = (AppCtx*) ctx;
>>>> PetscInt mybase, rowcol[] = {0,1,2};
>>>> const PetscScalar *x;
>>>>
>>>> PetscFunctionBeginUser;
>>>> ierr = MatGetOwnershipRange(B,&mybase,NULL);CHKERRQ(ierr);
>>>> ierr = VecGetArrayRead(X,&x);CHKERRQ(ierr);
>>>> if (mybase == 0) {
>>>> const PetscScalar J[] = {a + appctx->p1*x[1]/appctx->N, appctx->p1*x[0]/appctx->N, 0};
>>>> ierr = MatSetValues(B,1,&mybase,3,rowcol,J,INSERT_VALUES);CHKERRQ(ierr);
>>>> }
>>>> if (mybase == 1) {
>>>> const PetscScalar J[] = {- appctx->p1*x[1]/appctx->N, a - appctx->p1*x[0]/appctx->N + appctx->p2, 0};
>>>> ierr = MatSetValues(B,1,&mybase,3,rowcol,J,INSERT_VALUES);CHKERRQ(ierr);
>>>> }
>>>> if (mybase == 2) {
>>>> const PetscScalar J[] = {0, - appctx->p2, a};
>>>> ierr = MatSetValues(B,1,&mybase,3,rowcol,J,INSERT_VALUES);CHKERRQ(ierr);
>>>> }
>>>> ierr = VecRestoreArrayRead(X,&x);CHKERRQ(ierr);
>>>>
>>>> ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> if (A != B) {
>>>> ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> }
>>>> PetscFunctionReturn(0);
>>>> }
>>>>
>>>> This code does not provide the correct result, that is, the solution is the initial condition, either using implicit or explicit methods. Is the way I defined these objects wrong? How can I fix it?
>>>> I also tried to print the Jacobian with the following commands but it does not work (blank rows and error message). How should I print the Jacobian?
>>>>
>>>> ierr = TSGetIJacobian(ts,NULL,&K, NULL, NULL); CHKERRQ(ierr);
>>>> ierr = MatAssemblyBegin(K,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> ierr = MatAssemblyEnd(K,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
>>>> ierr = MatView(K,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
>>>>
>>>> I would very much appreciate any kind of help or advice.
>>>> Best,
>>>> Francesco
>>>>
>>>>> Il giorno 2 apr 2021, alle ore 04:45, Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> ha scritto:
>>>>>
>>>>>
>>>>>
>>>>>> On Apr 1, 2021, at 9:17 PM, Zhang, Hong via petsc-users <petsc-users at mcs.anl.gov <mailto:petsc-users at mcs.anl.gov>> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>> On Mar 31, 2021, at 2:53 AM, Francesco Brarda <brardafrancesco at gmail.com <mailto:brardafrancesco at gmail.com>> wrote:
>>>>>>>
>>>>>>> Hi everyone!
>>>>>>>
>>>>>>> I am trying to solve a system of 3 ODEs (a basic SIR model) with TS. Sequentially works pretty well, but I need to switch it into a parallel version.
>>>>>>> I started working with TS not very long time ago, there are few questions I’d like to share with you and if you have any advices I’d be happy to hear.
>>>>>>> First of all, do I need to use a DM object even if the model is only time dependent? All the examples I found were using that object for the other variable when solving PDEs.
>>>>>>
>>>>>> Are you considering SIR on a spatial domain? If so, you can parallelize your model in the spatial domain using DM. Splitting the three variables in the ODE among processors would not scale.
>>>>>
>>>>> Even though it will not scale and will deliver slower performance it is completely possible for you to solve the 3 variable problem using 3 MPI ranks. Or 10 mpi ranks. You would just create vectors/matrices with 1 degree of freedom for the first three ranks and no degrees of freedom for the later ranks. During your function evaluation (and Jacobian evaluation) for TS you will need to set up the appropriate communication to get the values you need on each rank to evaluate the parts of the function evaluation needed by that rank. This is true for parallelizing any computation.
>>>>>
>>>>> Barry
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>>
>>>>>> Hong (Mr.)
>>>>>>
>>>>>>> When I preallocate the space for the Jacobian matrix, is it better to decide the local or global space?
>>>>>>>
>>>>>>> Best,
>>>>>>> Francesco
>>>>
>>>>
>>>>
>>>> --
>>>> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
>>>> -- Norbert Wiener
>>>>
>>>> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
>>>
>>>
>>> --
>>> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
>>> -- Norbert Wiener
>>>
>>> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
>>
>>
>> --
>> Stefano
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210420/9846343a/attachment-0001.html>
More information about the petsc-users
mailing list