Parallel GPU solver for PLAXIS

Dear Amir,

PLAXIS does not utilize GPU yet. We only utilize the physical cores of the CPU to run parallel calculations and processes.

Here are some useful topics for you:

• https://communities.bentley.com/products/geotech-analysis/w/plaxis-soilvision-wiki/45854/parallel-calculations
• https://communities.bentley.com/products/geotech-analysis/w/plaxis-soilvision-wiki/45816/is-plaxis-3d-enabled-for-multicore

Dear

Many thanks for your information. I have a question.

My computer is the Octa Core (8) core computer.

• If I define for all my phases to only use 1 core per phase, I could run 8 independent phases at the same time in parallel. In this regard, could I choose Picos (multicore iterative) option or it is just compatible and workable with 2 or more cores per phase to be defined?

I mean could I choose the maximum 1 cores per phase to be used and at the same time select the Picos (multicore iterative) option to use the Picos ability?

• Which one is the fastest way to run and analyze the complex models. If I choose the maximum 1 core per phase to be used to run 8 independent phases at the same and select the option Picos-multicore iterative option (only if it is applicable for 1 core/phase) or If I choose the 8 core per phase and run just one phase at the time and leaving no room for additional phase calculation to be run at the same time and select the Picos (multicore iterative) option to use its ability?

I would be thankful if you may please guide me in this process

Thanks for your response

Sincerely

Amir Farshadfar

Dear Amir,

In principle, the power of PICOS comes from the fact that there is an iterative procedure and it benefits from solving the system of equations in parallel on multi-core processors. That is also explained in the article I shared before.

If you use PICOS with a single-core then you lose the multi-core solving of the equations but you keep the iterative procedure, which many times is better than the Classic solver.

Regarding the fastest way to run an analysis, this really depends on what you are analysing. Every project is different (2D or 3D), with lots of details, structural elements, etc.. It is difficult to answer whether one or the other will have a better performance unless we test your project.

In general, when using more cores, the calculation kernel will need to divide the calculation over the different processors (cores/CPUs). However, this comes with some costs: the calculation needs to be split, and then the results should be merged. After some testing, we found out that a small number of elements divided over too many cores actually becomes slower to calculate. For that reason, we have a threshold value for the number of dofs (degree of freedom, dof) per core, which is about 10,000 dofs. So, if this number becomes too low, we just use fewer cores not to slow down the calculation. 

For larger models, you will see an increase in the number of cores used.

In your case, you will need to test what is the optimum performance for that specific project, is it 2 cores per phase, so you compute 4 phases together or 4 cores per phase so you compute 2 phases at a time. 

Can we simulate the a 1 km  1.7 km  70 m large land sliding by the Plaxis 3D?

Can we simulate the 1 km * 1.7 km * 70 m large land sliding by the Plaxis 3D?

Dear Amir,

One should consider the number of elements that are required to be generated to be able to retrieve accurate results. In general, when doing geotechnical numerical analysis we should see how to have an efficient geometry and meshing in order to achieve an outcome we can interpret.

Therefore, the question is not whether it is possible to simulate a large landslide, it is a question of how accurate can you create a model that will give a response that you can accept with confidence.

Just to give you a simple example, if your shear failure surface occurs at X depth, in which the finite elements are 5m big, one can challenge whether this is captured properly and accurately.

I would start small first, test my model and then proceed to a model this big.

Dear Amir,

One should consider the number of elements that are required to be generated to be able to retrieve accurate results. In general, when doing geotechnical numerical analysis we should see how to have an efficient geometry and meshing in order to achieve an outcome we can interpret.

Therefore, the question is not whether it is possible to simulate a large landslide, it is a question of how accurate can you create a model that will give a response that you can accept with confidence.

Just to give you a simple example, if your shear failure surface occurs at X depth, in which the finite elements are 5m big, one can challenge whether this is captured properly and accurately.

I would start small first, test my model and then proceed to a model this big.