Re: Problems with scheduler

[David Trueman <david@cs.dal.ca>]

On Fri, 21 Apr 1995 brtmac@ksu.ksu.edu wrote:

> I'm curious about one aspect of zmailer's performance.  We have a
> system with >11,000 entries in the passwd file.  We are running
> sendmail 8.6.12 on a SS5/70.  sendmail is currently able to handle the
> load, but until we upgraded to the SS5 from an IPC stock sendmail
> wasn't able to.  I was able to make sendmail keep up with the load by
> replacing the getpw routines with routines that accessed a gdbm file
> with the passwd information stored in it.
> 
> How does zmailer access the passwd file?  Is anyone running zmailer on
> a system with a large passwd file?  If not, how do people deal with
> thousands of mailboxes?

I believe that by default zmailer scans the password file sequentially
using getpwent() unless you set GETPWNAM= in the hostenv file.

However, we use the Berkeley ?dbm routines for our passwd access so
sequential access is inappropriate.  When I set GETPWNAM= so that it now
use getpwnam() nad getpwuid() and therefore hashed access, the performance
improved *dramatically*.  I also made another change:  through a private 
version of the library routines, I turned off the retrieval of the
encrypted password that normally happens for a uid==0 process -- this also
made a big difference.

The two changes made the difference between router often consuming 50%
of the CPU (SS10/41) to now when it seldom consumes more than 10%,
presumably because it is now I/O-bound.  This is on a 7,000 line passwd
file.

  David Trueman,
    Systems Manager, Dalhousie Math, Stats and Computing Science
    Co-Chair, Metro*CAN Society running Chebucto Community Net