Parallel computing: How to use it in Flux?
Default settings
By default, the multithreading option, allowing the use of several cores for multithreaded algorithms (integration and assembly), is activated and half of cores of the machine is used. Distributed computing (MPI) requires a user configuration.
The settings of these features are available from Flux Supervisor in the menu .
User configuration
| Step | Action |
|---|---|
| 1 |
In Flux Supervisor … Click on Options Select the Parallel Computing tab |
| 2 | In the Multithreading area, specify the number of cores to be used for multithreaded algorithms (integration and assembly) |
| 3 |
In the Distributed computing (MPI) area, specify resources useable by the distributed MUMPS solver. Click on Add to add a first computation node and specify its name and the maximum number of processes which can be run on it. This number should be equal to the number of physical processors of the machine. Repeat the operation for each computation node. Modify if necessary the Total number of MPI processes: this number should be less or equal to the total number of processes specified in the machine section. Modify if necessary the Number of threads per MPI process: for multicore processors, it is recommended to use as many threads as cores per physical processor. |
| 4 | Click on Apply to save. |
Note
In the Multithreading area, Flux Supervisor is not able to make the difference between a real and a virtual core (due to hyper threading technology). For instance, 4 cores architecture with hyper threading gives 8 detected cores. Although these virtual cores do not deteriorate significantly the computation time, it is better not to use them in order to save numerical memory.
The Distributed computing (MPI) part is used to specify resources useable by MUMPS. It is necessary to specify this solver in the solving options of the project, by selecting either MUMPS: parallel direct solver (Local) or MUMPS: parallel direct solver (Distributed).