Optimal nevt_job for 1M events?

Asked by Richard Ruiz on 2018-02-08

Hi Folks,

Just a quick consultation question:

For a smaller setup of O(10-20) cores, is there an optimal value for nevt_job when generating O(500k-1M) events at NLO+PS? Does this rule of thumb scale for O(5) or O(500) cores? If this is process specific, how should one approach p p > t tbar vs p p > w+ w-?

Thanks for any suggestions that you may have, it is very much appreciated!

Best,
Richard

Question information

Language:
English Edit question
Status:
Solved
For:
MadGraph5_aMC@NLO Edit question
Assignee:
Rikkert Frederix Edit question
Solved by:
Richard Ruiz
Solved:
2018-02-09
Last query:
2018-02-09
Last reply:
2018-02-09
Rikkert Frederix (frederix) said : #1

Dear Richard,

There is not really a rule of thumb since it all depends on details on the system you are running on. In principle the overhead of splitting jobs is negligible from the MG5_aMC side when running at NLO+PS accuracy. Hence, the lower the nevt_job the better you are filling your cluster and, hence, the faster you get your results. However, depending on how much time it takes to copy files back and forth to the cluster nodes, the time it takes for jobs to start, etc., there can be quite a sizeable overhead from the system/cluster point of view. Therefore, having too quick jobs (i.e., with too few events) might actually be quite a bit slower. It really depends on your specific setup.

Best,
Rikkert

Richard Ruiz (rruiz) said : #2

Hi Rikkert,

Thanks for clarifying and helpful information!

Best