[Tfug] Punched cards and paper tape.

[JD Rogers]

> So at 4 o'clock if there were still a line of people doing COBOL I'd strike
> up a conversation with one and while they were distracted I'd cut in the
> lace card into they deck.   After a few times of this the COBOL folks
> stopped getting in late, so the FORTRAN could start at 4.

Hah, gaming the queue. So much has changed, and yet so little. I'm
running Monte Carlo ray tracing simulations on Northwestern's newly
upgraded cluster (7000 cores, 4-6 GB ram per core, infiniban network,
generally heaps of fun). It supports a lot of users and so there are
pretty severe limits on the queue, no more than 512 tasks on the
priority queue with a 4 hours time limit, 1024 on the normal queue,
etc.

Most people have to worry about their job geometry and scaling with
MPI, but MC is trivially parallel, so I can submit jobs up to the
limit of each queue. Since I am geometry independent, I can also
submit jobs with the exact geometry and time limit that fits into the
current available procs, so I rarely have to wait for my job to start.
I ran 120,000 cpu hours of simulation last month.

So I guess I'm kinda like the COBOL users? On the other hand, if I can
take the average core usage from 96% to 98%, that's just good resource
management, right? Now if I can debug the SIGINT propagation through
the scheduler and convince the admins to make a preemptable queue, I
could in theory soak up all idle cores and take it to 100%.
JDR