Example: suppose your application searches a parameter space defined by a single integer offset N. Calibrate how many iterations complete in one hour, then assign each job a distinct starting point:
Job 1: process range starting at 0
myapp --start 0 --count 10000
Job 2: process range starting at 10000
myapp --start 10000 --count 10000
Job 3: ...
myapp --start 20000 --count 10000
The --count value (iterations per job) should be determined so the expected wall time is approximately one hour. If the workload is uneven across the parameter space, err on the side of shorter jobs and submit more of them. |