Hi everyone,
I need to build an equipment to run a fem model on FEniCS. I have made a speed test on intel i5-4200H(2 cores with 3.8GHz) @ 8GB RAM, which is an unacceptable results (whole running progress will cost almost one year). Any advices for improving equipment? my budget is about 5000USD.
I also have some questions about FEniCS.
Thanks in advance
Have a look at FEniCS + MPI on docker inefficient?.
You won’t gain much from using more processors for accelerating each run.
But since you are asking about monte carlo runs elsewhere, buy a few cheap machines with many GHz and sufficient GB RAM. This is your best option. Alternatively, use somthing like Microsoft Azure or other compute clouds.
@Don, i think you fail to aknowledge that there are scaling tests for dolfin that show close to optimal rates, see:
Just because the particular demo you looked at did not show optimal convergence, i think it is quite ignorant to say that every code is like this.
There are many codes branching of from dolfin, gathering HPC specifically, like fenics-hpc.
For scaling results see one of their papers.
Similarly, the development version of dolfin, called dolfinx run performance tests nightly, see this link for results.
Don’t you agree that it’s more efficient to run multiple serial runs rather than distributing each run over multiple processes, provided that the problem is small enough to run serially?
Still, I am missing, where you read about “close to optimal rates” in this same paper.
I am not denying that it can be more efficient to distribute multiple serial runs if the problem is small enough, see for instance: How to achieve parallel running on FEniCS like parallel command in Matlab?
However, if the problem is large, there is gain in using multiple processors, see for instance figure 12 in that paper. Do you think the scaling in that figure is far of? The exact phrasing in the paper is:
Figure 12 presents strong scaling results for the steam turbine with over 36 M thermal and over 108 M displacement degrees-of-freedom (linear and quadratic elements).
The runtimes are good, and the scaling satisfactory, as observed for the turbocharger.