[klausb]

I am looking for a (boolean) trigger that tells me whether the matrix has changed since the last check to be able to cache and re-use data.


The idea is to toggle between true and false and re-use data as often as possible to avoid expensive re-calculations.

Is there something like that?

[amirhosseintaran]

Originally Posted by klausb

I am looking for a (boolean) trigger that tells me whether the matrix has changed since the last check to be able to cache and re-use data.


The idea is to toggle between true and false and re-use data as often as possible to avoid expensive re-calculations.

Is there something like that?

Hi Klaus,

I don't understand why you can't define this boolean yourself.

e.g. comparing the matrix with itself shows whether it has changed or not!


Regards,
Amir.

[jczhang]

Originally Posted by amirhosseintaran

Hi Klaus,

I don't understand why you can't define this boolean yourself.

e.g. comparing the matrix with itself shows whether it has changed or not!


Regards,
Amir.

Wow, I also had the same question. To compare, you need to save the old matrix. There is a cost in saving the matrix and then comparing. We want to avoid the cost.

[klausb]

The point is indeed, that comparing entire matrices has a significant impact on data storage requirements and overall performance even when using hash algorithms and I am looking for ways to speed simulations up.

[arjun]

Originally Posted by klausb

The point is indeed, that comparing entire matrices has a significant impact on data storage requirements and overall performance even when using hash algorithms and I am looking for ways to speed simulations up.

Just save the sum of diagonals and sum of off diagonals in two variables. Just compare these two values and see if they changed above a thresold.

You can also add, min, max of dia and min max of off diagonals. This was you do not need to store last matrix.

[jczhang]

I feel all these approach are unsafe with respect to false positive, i.e., the sums of entries of two matrices are equal but the two matrices are not equal.

[arjun]

Originally Posted by jczhang

I feel all these approach are unsafe with respect to false positive, i.e., the sums of entries of two matrices are equal but the two matrices are not equal.

Only way to be hundred percent sure is to cache the matrix. Any other method will always have some doubts.

[jczhang]

Originally Posted by arjun

Only way to be hundred percent sure is to cache the matrix. Any other method will always have some doubts.

It depends on the design. For example, in petsc a matrix has a two integer attributes: state and nonzerostate. Whenever the sparsity pattern changed, the nonzerostate is increased; and whenever an operation modifying matrix values is called, the state is increased. With that, one can easily and safely detect matrix changes.

[dlahaye]

Originally Posted by klausb

I am looking for a (boolean) trigger that tells me whether the matrix has changed since the last check to be able to cache and re-use data.


The idea is to toggle between true and false and re-use data as often as possible to avoid expensive re-calculations.

Is there something like that?

1/ I think it is save to assume that the matrix changes at every inner iteration of e.g. a SIMPLE iteration. You might this be looking for deviations from this rule;

2/ When is matrix assembly a computational bottleneck?

[klausb]

@dlahaye


1/ I think it is save to assume that the matrix changes at every inner iteration of e.g. a SIMPLE iteration. You might this be looking for deviations from this rule;

>> Not really, it depends on the simulation type, e.g. moving mesh...


2/ When is matrix assembly a computational bottleneck?

>> It's not. A matrix change, changes the eigenvalues, which triggers computationally expensive re-calculations which I want to keep to a minimum

[arjun]

Originally Posted by jczhang

It depends on the design. For example, in petsc a matrix has a two integer attributes: state and nonzerostate. Whenever the sparsity pattern changed, the nonzerostate is increased; and whenever an operation modifying matrix values is called, the state is increased. With that, one can easily and safely detect matrix changes.

Then i think you do not understand the problem that has been discussed.

Every iteration an operation modiying the matrix IS performed thanks to descretization.

What we are talking about is matrix being same as last iteration that one obtained after descretization. This matrix will stop to be different once the convergence is achieved.