Clarification on TOTALJOBS / MAXWAITINGJOBS
We need clearer documentation about these two very important parameters. The current one is insufficient, and in fact, it doesn't consider the changes in v3.13.0 (since this version, wrapped jobs count as one job, instead of the sum of wrapped and non-wrapped tasks).
Taking this into account, I understand that:
- TOTALJOBS: Number of active (queuing, submitted, unknown, running) jobs. We understand a job as any batch or interactive job submitted or executed in a system containing one or more (wrapped) individual tasks.
- MAXWAITINGJOBS: Number of queuing and submitted individual tasks. Wrappers increment this value in as many units as individual tasks have inside.
To support @jescriba in executing his experiments, I did a few tests (dummy exp a52r
).
- TOTALJOBS = 60, MAXWAITINGJOBS = 10, MIN_WRAPPED = 2, MAX_WRAPPED = 200
Wrappers have 10 tasks because MAXWAITINGJOBS = 10.
- TOTALJOBS = 60, MAXWAITINGJOBS = 15, MIN_WRAPPED = 2, MAX_WRAPPED = 200
Wrappers have 15 tasks because MAXWAITINGJOBS = 15
- TOTALJOBS = 60, MAXWAITINGJOBS = 12, MIN_WRAPPED = 15, MAX_WRAPPED = 200
Conversely, if MIN_WRAPPED > MAXWAITINGJOBS, Autosubmit cannot wrap 15 tasks that will instead be submitted individually. However, the last 12 tasks are wrapped. I guess because there cannot be 15 tasks to wrap. POLICY
is not set, so I guess it's FLEXIBLE
. In any case, this is strange behavior, and I think the same limit should apply to each member's first and last tasks.
At the running time, I tested different TOTALJOBS values:
- TOTALJOBS = 60: 6 wrappers (90 total tasks) are queued concurrently. This is expected because each wrapper is 1 job.
- TOTALJOBS = 2: 2 wrappers are submitted, and also 2 independent tasks. I didn't expect this behavior because these are 4 jobs. It' seems that's like wrappers, and individual tasks had independent counts (then I realized that this is due to the TOTAL_JOBS value, see below).
On the other hand, @jescriba reports that TOTAL_JOBS (defined at the platform level) has an impact in the wrapper sizes (like it didn't count jobs but tasks, like MAXWAITINGJOBS).