More 4.0 bashing

[William LeFebvre]

Just today I had to wait 50 seconds (at least) while my 4Meg 3/60 (running
4.0.1) took a coffee break.  What was I doing?  Well, in one window I had
emacs do a reverse search in a 2.3 megabyte file for a string that
(unintentionally) didn't exist.  I figured that might take a while, so I
moved the mouse over to a shelltool and typed a command.  The window was
not even *selected* (as in, no solid cursor and no highlighted border) for
at least FIFTY SECONDS!  The only things that moved on my screen were the
perfmeters and the clock and the mouse pointer (that's how I knew it was
50 seconds).  This never happened under 3.2.

To be fair, I should say that my client is being served off of titan, a
3/280 that is primarily used as a time-sharing machine.  It typically has
25 to 35 people logged directly on to it banging away on it.  But on the
flip side, I have noticed that my 3/60 will go catatonic---completely lock
up to the point where SunView isn't changing active windows with the
mouse---for 20 to 30 seconds at a time.  While this is happening, I look
at my perfmeters.  My client is doing nothing, I see very light paging and
swapping activity, even the collision meter is quiet.  And, most
importantly, my server is at 50% CPU or less.  So WHY is it out to lunch?
Who knows?

I must say that I am not completely satisfied with 4.0 performance.  I
went into this with an open mind, but it is noticeably slower than 3.2.
Note that I did not say substantially slower.  But it is noticeable.  We
are running a tailored kernel.  We have turned off alot of the daemons.
We have done many of the tuning things that Sun recommends.  I speculate
that what's killing my machine is weird NFS paging or swapping activity.

One more thing:  has anyone else noticed that a "df" on a 4.0.1 NFS server
takes a ridiculously long amount of time?  Titan has two xy drives with 8
partitions mounted between them.  It also has 6 NFS partitions mounted.
Usually "df" will take between 5 and 10 seconds before even the initial
banner is displayed.  My client doesn't exhibit this behavior.  This is
very odd.

	William LeFebvre