what's this "nice" unix thing? I'm a sysadmin...

Eating Before Swimming (mathboy@sizone.org)
Fri, 23 May 1997 10:41:19 -0400 (EDT)


then seth's all..

> understand firewalls well enough. As for smart.net, I tried running the
> client on my shell account at the ISP, but the sysadmin threw a fit and
> in no uncertain terms told me not to run that client again. I sometimes

I have got this EVERYWHERE I try to run or try to get friends to
run clients.

It seems that there needs to be some updated version of xload or the
kernel's "load avg" facility. Or admins have to be educated a lot more.
Or they have to stop trying to charge dollars for cycles.

Strange, all this, since Im not a 'seasoned' admin who deals with users
doing all sorts of nastyness to the system, since they deal with CS
students and the like. (I usually scare my clients into not messing
with the stuff we setup on their boxes, and they're usually good about it. :)

These admins dont seem to know what "niced" means. Yes, they may not
care in some cases, there may be a policy of not running 'external
code', but I am RAILING against those cases where the admin
specifically said "your process is getting 60% of the cpu. its too much.
we killed it. Look, it has over 1hr of cumulative CPU!"

And its not like they said "that created over 200 kilo joules of extra
heat and wear on the components and taxing our air conditioning", which
I might vaguely accept as a logical excuse. ;)

It ONLY gets 60% of the cpu when there is 60% LEFT OVER after everything
else has got its share (unless its niced even heavier than DESCHALL,
and I've been using -19 on most systems I get onto for it -- on one
loaded ultraSparc 2CPU I was getting ~0.05% and ~0.04% CPU for over 8 hours
before everyone went home).

THis will bring the load average right up to 1.00 or very close to it.
DESCHALL uses what it can get, and not more, and uses none when
everyone else is busy. You'd think this would RING SOME BELLS in their
heads: "Curious, that, this load average of 1.00 all the time. Hmmm."
But no. They're also allowing other processes running at 0 to eat
up alot more CPU than that, which are named 'netrek' and the like.

Xload needs to have a blue section at the bottom of its new colour coded
load display: blue for NICE, and nasty RED for normal usage from huge
monstrosities like NETSCAPE, which I saw one student running over 5
runaway copies. (And then there was the LPD, which musta been running
a SolNet client inside it or something ;), using up 87% of the cpu on
a 4 processor ultraSparc with an empty LPQ...)

And TOP(1) on most of the boxes doesnt show the educational "Niced: 50.0%"
stat either. This is ignored on those that do have it as well.

I've even taken to renaming my damn client "very-niced-des-challenge-sparc"
and the like. They responded with "well, it must not be very niced, cuz
it got xx% of the CPU for several hours." Like, duh.

Actually, the colour coded XLOAD wont work - they'll see the 1.0 load
and STILL freak. What needs to happen is "ALOAD" needs to be written -
"Available CPU meter". You set it to display how much CPU is available
at 0 nice level - a rolling average similar to current load stats: just
subtract and completely IGNORE anything which is niced. Thats it.

I think all this (xload and hysteria about ANY cpu being used) stems
from the days when CPU usage was charged per cycle in hard currency to
research groups and the like... $0.07 cents per CPU hour, $0.04 niced...

Sigh.

Anyone feel like writing up a very formal, polite and educational
plea/explanation as to HOW niced processes work and why its ok to use left
over CPU for DESCHALL on systems where you ARE allowed to run 'external'
programs? This has happened to me on no less than 12 systems
I've or got friends to run this on.

/kc