|
[Sponsors] |
Runs on multiple nodes are 2x slower than on one node |
![]() |
|
LinkBack | Thread Tools | Search this Thread | Display Modes |
![]() |
![]() |
#1 |
New Member
Ivan Pombo
Join Date: Jan 2024
Posts: 1
Rep Power: 0 ![]() |
Dear All,
We have been testing locally with the usual Github installation of Reef3D in Linux machines. We have been using one of the Flow around a Circular pier tutorial with the CFD solver to test some runs with different computational configurations. We have tested running this simulation in two different settings: - single node: single Intel Xeon machine with 30 CPU cores. - two nodes: two Intel Xeon machines, each with 30 CPU cores, summing up to 60 cores. Running the same simulation on each setting, modifying only the M 10 parameter, we are obtaining an absurd computational time with the setting composed of two nodes. In practice, the simulation on a single node took 1h30mins and on the two nodes is taking more than four hours. While we expected it to not be faster, we didn't expected it to be significantly slower. So we are doubting our installation procedure is somehow wrong to deal with simulations running on multiple nodes. To clarify, we have tried with smaller machines and different Reef3D simulations and obtained the same behavior. Moreover, we have tested the cluster with OpenFOAM and everything seems to be working as expected. We are using apptainer to launch the mpi job from outside the apptainer and to all the machines, and have an NFS server in one of them. Any help is appreciated to figure this out! Thanks in advance!! |
|
![]() |
![]() |
![]() |
![]() |
#2 |
Member
Alexander Hanke
Join Date: Dec 2023
Location: Trondheim
Posts: 39
Rep Power: 3 ![]() |
Each case has a number of cells per partition under which the CFD calculations take less time than the communication between partitions. CPU communication speed is usually in this order: CCD > CPU > Socket >>> other computer
So you can either just use one CPU or even only parts of one (e.g. SLURM allows partial usage of a node) or increase the number of cells by decreasing the grid spacing if you want to fully utilise your machines. Hope that helps you. |
|
![]() |
![]() |
![]() |
![]() |
#3 |
Super Moderator
Hans Bihs
Join Date: Jun 2009
Location: Trondheim, Norway
Posts: 396
Rep Power: 18 ![]() |
Hi Ivan,
the tutorial cases have quite coarse resolution, so possibly the total cell count is too low to assure good scaling. Can you try with a large number of cells (try ca. 10 000 cells per core at a minimum). How many cells are you running OpenFOAM at? |
|
![]() |
![]() |
![]() |
![]() |
#4 |
Member
Felix S.
Join Date: Feb 2021
Location: Germany, Braunschweig
Posts: 88
Rep Power: 6 ![]() |
Hey there,
I just wanted to add something I experienced on one HPC I ran REEF3D on. Running REEF3D with OpenMPI compiled by gcc resulted in bad scaling similar to your findings. However, after compiling REEF3D and hypre with icc and intelmpi scaling was just fine. Maybe recompiling with a different compiler might therefore solve your problem? Good luck anyway! Last edited by Felix_Sp; February 7, 2024 at 12:43. |
|
![]() |
![]() |
![]() |
Tags |
mpi, multiple nodes, reef3d |
|
|
![]() |
||||
Thread | Thread Starter | Forum | Replies | Last Post |
Openfoam running extremely slowly using multiple nodes on HPC | wdx_cfd | OpenFOAM Running, Solving & CFD | 3 | September 19, 2023 05:17 |
Hybrid OpenmMP+MPI optimisation | geronimo_750 | SU2 | 8 | April 6, 2022 07:29 |
[ICEM] Error in mesh writing | helios | ANSYS Meshing & Geometry | 21 | August 19, 2021 14:18 |
problem with interfoam on multiple nodes | gkara | OpenFOAM Running, Solving & CFD | 1 | July 6, 2016 09:17 |
CFX4.3 -build analysis form | Chie Min | CFX | 5 | July 12, 2001 23:19 |