From JKONG@ac.dal.ca  Thu Sep  9 16:42:03 1993
Date: Thu, 09 Sep 1993 19:42:03 -0300
From: JKONG@ac.dal.ca
Subject: Summary: queuing system and priority
To: chemistry@ccl.net
Message-Id: <01H2R8FP6WTE00L65E@AC.DAL.CA>


Dear netters,

	Last month, I sent several queries about queuing system
and nice command on ibm rs/6000. Here I just give the summary
of replies and what I learned from them. I hope this message would be
useful to those who just got machines and are going to install NQS and
NFS on them, especially ibm rs6k.

	When we got the two ibm rs6k's, there were several problems:

(1) The original printer queues do not work very well. They do not have
the features you want; they are just queues. IBM has NQS but we were asked
$2,000 for it.

(2) Some people are willing to run at low priority because their jobs are
long and cpu-bound. But the "nice" doesn't work well. I got several replies 
which basically have the same complain. A recent message in this net 
titled "nice on AIX" states this problem very well.

(3) Interactive processes(commands) are slow if there are a couple of jobs 
running.

(4) Only two(!) loggin sessions come with each machine and again we have to 
pay for more($500 each).


	Following the suggestiong of Besler(bbesler@ouchem.chem.oakland), 
Max(mholtha@neon.chem.tu-berlin.be) and several other people, we 
installed DQS, got from anon ftp site ftp.scri.fsu.edu. Thanks to Martin's
(schuetz@ips.id.ethz.ch) suggestion, I changed the system call setpriority 
in subroutine setlimits, which sets nice value, to setprio which set a fixed 
priority for the calling process. Now a job in a low priority queue always 
sits idle whenever there is a job in a high priority queue running in cpu. 
Here is a tip: do not set a priority to 127(lowest possible) because a 
kernal process kproc has a priority of 127. Actually, the cpu time kproc 
consumes(ps -Af will show) tells you exactly how much time the cpu has sit 
idle. I set priorities of queues from 126 to 124. This makes interactive 
processes always run at a way higher priority than queued jobs. No obvious 
slow-down for running commands is found on our 580 even when 3 jobs are 
running. This arrangement overcomed the problem 1, 2 and 3. With the 
combination of NFS and DQS the requests for loggin sessions are spreaded out 
and one loggin for one user in a cluster should be good enough in most time.

	Finnally, DQS is easy to install and fun to use.

I am grateful to all the guys who replied to me!

Jing