performance problems caused by login pile-ups

From: Padraig Houlahan <houlahap_at_ucs.orst.edu>
Date: Wed, 15 Jan 1997 15:56:18 -0800 (PST)

I hope someone can help with this...

We run a 2100 alpha with 3 CPUs and 512 MB of RAM as an academic mainframe.
We also run DU4.0 with C2 security.

All faculty, staff, and students have an account so that we might have about
250 - 290 users online in the afternoons..

Our problem is that in this university environment, we get surges caused
by users trying to login right after class ends, and so on mondays and
wednesdays we might have more than 60 logins within a few minutes
of the top of the hour. As a result, our load averages can climb way
above 100 (getting up to 160+ occurred on monday) and take 10 to 15 minutes
to drift back down.

(We keep performance charts showing loads and users at www.orst.edu/~ucs and
these show what happens fairly well)

We have tried blocking logins through creating a /etc/nologin file to allow
the system to catch up, but this doesn't work - probably because other processes
are started before the nologin file is checked for.

My Questions:

1. Has anyone encountered this kind of performance problem and successfully
        managed it?

2. Is there a safe way that could be implemented by a shell script, to send
        sleep (or stop) signals to an appropriate process to stop all login
        type processes and then restart them once the load average has
        recovered? If so, which processes would be the appropriate ones to
        control.

TIA

PH

-- 
Padraig Houlahan                Computer Services - IS
houlahap_at_ucs.orst.edu		Milne Computer Center - 217            
503-737-0671 (W)		Oregon State University        
				Corvallis, OR 97331-5202              
Received on Thu Jan 16 1997 - 01:07:36 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:47 NZDT