Hello everyone,
I am still new in running job using code saturne and I have a question on running job in parallel. Since there is no step-by-step guide for running the job in parallel for v4 and this is the first version that I'm working with. I wonder if there's anyone who could guide me on how to start to run the job in parallel. FYI, i'm trying to run my jobs on university cluster. Thank you.
Cheers,
Muhsin
Run Code Saturne 4 in parallel
Forum rules
Please read the forum usage recommendations before posting.
Please read the forum usage recommendations before posting.
-
Yvan Fournier
- Posts: 4208
- Joined: Mon Feb 20, 2012 3:25 pm
Re: Run Code Saturne 4 in parallel
Hello,
Did you install the code, or did somebody else install it ? On a cluster, it is important to do the "post-install" of the code to tell it what batch system is used.
After that, running in parallel (using setups from the GUI) is almost the same as running in serial (choosing the number of processes in the runcase, preferably using the GUI if you are not accustomed to running in parallel).
User subroutines require more precautions, but there are examples dedicated to this.
Also, performance is not always "linear" with the number of processes, so you need to find the optimum for your case (starting from 50000 cells/process if possible), but that can be done after a first run on a few processes.
Regards,
Yvan
Did you install the code, or did somebody else install it ? On a cluster, it is important to do the "post-install" of the code to tell it what batch system is used.
After that, running in parallel (using setups from the GUI) is almost the same as running in serial (choosing the number of processes in the runcase, preferably using the GUI if you are not accustomed to running in parallel).
User subroutines require more precautions, but there are examples dedicated to this.
Also, performance is not always "linear" with the number of processes, so you need to find the optimum for your case (starting from 50000 cells/process if possible), but that can be done after a first run on a few processes.
Regards,
Yvan
Re: Run Code Saturne 4 in parallel
Hello Yvan,
Thank you for your reply. It's a good start for me. Yes someone has installed it for me, "post-install" also has been done.
It's running but I'm raising the question because it seems to me it's running quite slow at this moment, especially when it writes files. I think it's working, just to like reconfirm about it. Maybe it's like what you've said, I need to find the optimum performance for my case. Here I attached the runcase file that I select from GUI, in case you would like to check the one that I use. No additional subroutines provided at the moment everything is set from the GUI.
Cheers,
Muhsin
Thank you for your reply. It's a good start for me. Yes someone has installed it for me, "post-install" also has been done.
It's running but I'm raising the question because it seems to me it's running quite slow at this moment, especially when it writes files. I think it's working, just to like reconfirm about it. Maybe it's like what you've said, I need to find the optimum performance for my case. Here I attached the runcase file that I select from GUI, in case you would like to check the one that I use. No additional subroutines provided at the moment everything is set from the GUI.
Cheers,
Muhsin
- Attachments
-
- runcase.tar.gz
- runcase file
- (630 Bytes) Downloaded 208 times
-
Yvan Fournier
- Posts: 4208
- Joined: Mon Feb 20, 2012 3:25 pm
Re: Run Code Saturne 4 in parallel
Hello,
It looks OK (I'm really not a big fan of SGE, as we can't help automate the choice of "pe" for you due to its syntax, but I assume your administrator or colleague did that OK for you). I assume if you're not familiar with clusters, you administrator informed you you need to run "qsub runcase" instead of directly "runcase" to run in the parallel compute environment (but checking just in case).
How does the performance on a small number of ranks compare to the performance on a single node ?
By the name of the "pe" you use in GridEngine, it seems your cluster uses OpenMPI + Infiniband (which is a very common case). There are many variants, but if your performance is bad, there may be some OpenMPI "mca" parameters which you can define as environment variables which may help with tuning MPI parameters for Code_Saturne. Unfortunately, I'm out of office for the next two weeks, and will be at a conference after that, so I won't have access to our clusters (on which I could check our parameters) immediately.
Regards,
Yvan
It looks OK (I'm really not a big fan of SGE, as we can't help automate the choice of "pe" for you due to its syntax, but I assume your administrator or colleague did that OK for you). I assume if you're not familiar with clusters, you administrator informed you you need to run "qsub runcase" instead of directly "runcase" to run in the parallel compute environment (but checking just in case).
How does the performance on a small number of ranks compare to the performance on a single node ?
By the name of the "pe" you use in GridEngine, it seems your cluster uses OpenMPI + Infiniband (which is a very common case). There are many variants, but if your performance is bad, there may be some OpenMPI "mca" parameters which you can define as environment variables which may help with tuning MPI parameters for Code_Saturne. Unfortunately, I'm out of office for the next two weeks, and will be at a conference after that, so I won't have access to our clusters (on which I could check our parameters) immediately.
Regards,
Yvan