With Vuori (vuori.csc.fi) the users are able to run parallel jobs up to 144 processor cores without restrictions. However, bigger parallel jobs than 144 cores require that scalability tests are passed acceptably. One can apply for a temporary scalability permit from address helpdesk@csc.fi. The user has one calendar month time to run the tests and send the results back to scalability support.
The scalability tests are run by using the srun comand when MVAPICH2 is used and by using the mpirun command when OpenMPI is used and they should be run though the SLURM batch job scheduler using a batch job file submitted by the sbatch command. For more details and for examples of batch job scripts see:
The program runtime is detected with the time command (which is used in batch job file):
MVAPICH2:
time srun ./mpitest
OpenMPI:
time mpirun -np N ./mpitest
where N is the number of cores reserved.
The time command is needed for scalability tests only. In Vuori, the time command writes the timing statistics to "system error message output" file, which is defined in a batch job file with option: #SBATCH -e
To provide proper data about the scalability of the program, do the following steps:
- Generate one or more test runs which correspond your production runs well, but are clearly shorter in time.
- Run these batch jobs with Vuori using 8, 16, 32, etc., processors up to the wanted number of processors.
- Collect the timing statistics for an array of results.
- Send the data by email to CSC to e-mail address: helpdesk@csc.fi. (Note: Your message is not necessarily responded, unless we need more details.) Please mention, if your results are applicable for other users.
A parallel code has been, in general, judged at CSC to scale well when the speed-up factor is at least 1.5 when the number of processors has been doubled. CSC is also able to provide assistance in the code optimization problems, if needed.
Any possible questions can be sent to the e-mail address:
