Radioss solving
Altair EmployeeAnswers
-
Selva,
I believe it can be done. It's called cycle harvesting. Please get in touch with the PBS folks at Altair (pbssupport@altair.com) and someone will help you out.
0 -
Pradeep,
Thanks a lot.
I will try tat n revert back.
0 -
Hi Radioss users,
I want to give a fire for my crash analysis using Radioss solver.
I jst want 2 combine 4 machines for giving the run (Windows XP, 32bit processor)
Can any help me out in this issue...
regards,
Selva
First of all, you need these installations to do an spmd solve process with multiple hosts, all of them must be present on every host
1- HWSolvers Update (The latest one)
2- RDSPMD update (just extract the files into hwsolvers directory except radflex10_wnt)
3- MS-MPI
Then use the below code as a *.bat file, but before in your starter file set the /SPMD card (set the number of parallel process), this bat file must be in the same directory with your starter and engine file
set LM_LICENSE_FILE=7788@your_license_server
set ALTAIR_HOME=C:\Altairwin64\hw10.0\hwsolvers\bin\win64 (or win32)
COPY %ALTAIR_HOME%\radflex10_wnt.exe
%ALTAIR_HOME%\s10_p4win64_spmd.exe -input your_starter_file_0000.rad
MPIEXEC -hosts n computer_name_1 process_count_1 computer_name_2 process_count_2 ... %ALTAIR_HOME%\e10_p4win64_spmd.exe -input your_engine_file_0001.rad (n is the host number)
Example:
My /SPMD Card in the starter file
/SPMD
0 48 0
and MPIEXEC line then should be like that:
MPIEXEC -hosts 3 computer_1 16 computer_2 16 computer_3 16 ......... so with 3 hosts and 16 process in each host begins
Remeber that, your network connection must be fast enough and your bandwidth must be large enough for file transfer.
If your bandwidh is low then the limiting factor will be the network load, not the CPU's .
0 -
Hello !
Could You describe step by step how to run the Radioss calculations on the other computers in the network ?
I don´t really get what is about starter and engine files.
Up to now I was calculating on one machine by generating *.fem file from HyperMesh and using preinstalled Radioss solver to run the calculations with the *.fem file.
I would be very grateful for Yours help.
Regards
0 -
Hello !
Could You describe step by step how to run the Radioss calculations on the other computers in the network ?
I don´t really get what is about starter and engine files.
Up to now I was calculating on one machine by generating *.fem file from HyperMesh and using preinstalled Radioss solver to run the calculations with the *.fem file.
I would be very grateful for Yours help.
Regards
There are two known methods for parallel computing: Shared Memory Parallel (SMP) and Single Program Multiple Data (SPMD). You could check the meanings of them from the Radioss manuals.
When you create a 'block data' model for radioss, then you have two files (starter file 'example_0000.rad' and engine file 'examples_0001.rad'. Starter file contains your model, engine file defines how to solve it. Also check them from the manuals.
The above script that i wrote is for parallel computing of block data in a multicpu (and/or multihost) environment. You could use smp in one computer with multiple cores or you could use spmd with multiple cores and/or with multiple hosts.
For bulk data computing (example.fem) you have 1 file for both model and solution. Parallelism of that is a bit different but so similar to block data parallelism.
You need the same applications mentioned before (solvers, solver updates, spmd updates for hw10, mpi software (ms, platform, intel) ).
If all of them are installed correctly, you need a script file for hw10, not necessary but useful in hw11 also (it has its own control panel for that, you could also use it).
So please check out the above answer and modify it for your own purposes.
When you need to solve your model with multiple hosts you need that command:
N is the host count and N1, N2 is the number of processes for each host,
mpiexec -hosts N host1 N1 host2 N2 ... your_executable other_options
When you need to solve your model with multiple cpus in the same host you need that command:
mpiexec -N your_executable other_options
Pls also PM to me for your further questions about parallel solving.
Final Note: You need a good interconnect for your multiple host calculations due to the latency bottleneck.
0