ajacquey/ddmfrictionalslip.jl

A Displacement Discontinuity Method (DDM) implementation for fault slip

Hi there 👋

 

ddmfrictionalslip.jl's Issues

File path for outputs points to package installation folder instead of current script folder

The output files are found in the package installation folder instead of the folder where the script is.

Make us of collocation matrix properties to fasten assembly

The collocation matrix has a few properties (symmetric if order 0 as an example).
Should make use of these properties to fasten the assembly.

Multhreading and solve

The goal of this issue is to enable multithreading for the linear solver.
I found out that when using sparse matrices, IterativeSolvers.jl is not parallelized because of sparse matrix/vector multiplication.

I have several options here:

• Implement custom _threaded_mul! function using Threads. This could be challenging as the Jacobian is not symmetric for PWLC and PWQC (for PWC, it is symmetric). Here is a solution for symmetric matrices: JuliaLinearAlgebra/IterativeSolvers.jl#203
• Using MKLSparse.jl. Apparently, loading this package would allow multithreaded sparse matrix/vector multiplications.
• Switch back to dense matrices. I do not gain any memory by using a sparse matrix for this system (the sparse matrix actually takes a bit more storage than the dense one).

Avoid rebuilding jacobian

Currently, we reallocate the Jacobian after initializing the problem (because we have the correct number of dofs after initializing).
For large system, this can take a large amount of time!

Fix:
Update n_dofs in the addVariable! function everytime.
Just check value of n_dofs during the initializing.

Explore the use of CSV.jl for the output system

CSV.jl seems to be compatible with multithreading when reading/writing csv files.
Should investigate if replacing DelimitedFiles.jl with CSV.jl would make sens.

Extend to multiple variables

Extend the structure of the code for solving multiple coupled variables.

Simplify assembly

Currently, I am rebuilding the entire collocation matrix for each non-linear iteration.

In practice only the diagonal components of the collocation matrix are modified throughout the simulation.
To speed-up the simulations, one should:

• Build the collocation matrix only once before starting the simulation (in the solver constructor)
• Only recompute the diagonal elements (element-wise) of the collocation matrix during the assembly