Hello all,
I write to this list in order to see if some can provide some
clarification about what I see as an inconsistency between reservations,
accounting, and the output provided by sreport(most probably I’m missing
something).
I define my test:
* We are testing SLURM in a small test cluster: 2 nodes with 16 CPU
each (32 in total).
* Two users (user1 and user2), each one of them in a different account
(account1 and account2).
* We create a reservation of 10 minutes for the user1. In total it
should be 320 min/cpu:
scontrol create reservation starttime=2018-11-22T13:05 duration=10
user=user1 nodes=all
* The user1 does not submit any job, but the reservation is billed
anyways, as it should. As result, the command sreport cluster
displays 320 as the time consumed for user1 and account1. So far so
good.
~]# sreport cluster UserUtilizationByAccount start=2018-11-22
end=2018-11-22T23:59
Cluster Login Account Used Energy
--------- --------- --------------- --------------- -------- --------
test_cl user1 Account1 320 0
* Now the user2 submits a job which takes around 190 sec in 8 cpu,
which is around 25 cpu minutes. Again, everything’s good:
~]# sreport cluster UserUtilizationByAccount start=2018-11-22
end=2018-11-22T23:59
Cluster Login Account Used Energy
--------- --------- --------------- --------------- -------- --------
test_cl user1 Account1 320 0
test_cl user2 Account2 25 0
* Now we create a new reservation for the user2, and the user2 submits
the same job but inside the reservation:
~]# scontrol create reservation starttime=2018-11-22T16:05 duration=10
user=user2 nodes=all
~]& srun --reservation=user2_1 …
* And then, here it comes the inconsistency. These are the sreport
results:
sreport cluster UserUtilizationByAccount start=2018-11-22
end=2018-11-22T23:59
Cluster Login Account Used Energy
--------- ------------------------ -------- --------
test_cl user1 Account1 320 0
test_cl user2 Account2 269 0
Which is the result of the total reservation minus the two jobs
submitted (320- 25-25 and some rounding up). Corroborated several times.
But in theory, it should be around 345 (the whole reservation, no matter
if used or not, and the job submitted without reservation).
Am I misunderstanding something about the way slurm handles the
reservations?
Can anyone provide some insight?
Apologies for the long post, but I wanted to provide a consistent context.
Thanks for your time.
Alberto,
--
------------------------------------------------------------
Alberto García Fernández.
Ingénieur de recherche
CNRS - IDRIS
Tel: +33 (0)1 69 35 84 37
Mob: +33 (0)7 88 55 80 49 +34 669 692 272
Web: https://es.linkedin.com/in/alberto-garcia-fernandez-bb94449a
