[slurm-users] Differences between sacct and sreport

[Daniel Tosello]

Hey all,

The short version of my query is like: "how do i explain the
differences between the outputs of similar queries in sacct and
sreport to my management team?"

The longer version is, i'm guessing, something many of you could
contribute to... though i've searched the list and i haven't seen this
question asked here in a while, i know other people have expressed
concern with how different the output of, say, `sreport
AccountUtilizationByUser` and `sacct
--format=JobID%14,Account,User%30,CPUTimeRAW -apX | awk -F'|' '{a[$2 "
" $3] += $4} END{for (i in a) print i, "- " a[i]}'` are, but also that
they don't appear to have a consistent relationship.

Ultimately i'd like to have a description of how the results for each
query in sreport are generated, and how i can replicate them with
sacct. For us this is only really an operational problem because our
legacy figures all use sreport (which was simpler initially for
management) and we'd prefer to use sacct going forward (for more
specialized queries, now that people in management are getting used to
operations at scale).

>From browsing the code of each it kinda seems like sreport uses a
bunch of pregenerated indexes and sacct generates aggregate tables
while executing... does that sound correct? I suspect i'm going to
have to dive a bit deeper to get answers, but if anyone on the list
has them already that would be awesome.

Of course, if any of the Sched team want to explain why we see such
wild variations in some cases, or point me to docs dealing with it,
that would also be cool.
--Dan