After my soapbox-mode posting on the problems with malloc()/exec()/fork()
and the contiguous space issues, WHAM! It bites me back...
 
This time it hit my 960 which has only 2 system volumes, both 2Gb SCSIs.
Same characteristics as before, when largest contiguous space dwindled
down to ~256 sectors everything in Posix-land started failing with the
old "resource busy, try again" errors.  I called the HPRC and this time
didn't settle for the textbook CONTIGVOL answer.  After some experimenting
we discovered something "different".
 
I have always defined my disc volumes perm/trans space as 100/100 except
for ldev 1 which I set at 75/75.  This resulted in a pretty nice balance
of space allocation.  However, as many of you have noticed, after you go
on 5.0 Push, things start disappearing from ldev 1 and moving elsewhere.
Furthermore, it tries not to allocate anything on ldev 1 unless there is
no space anywhere else.  As a result, I had ldev #1 with about 100K sectors
of transient space in use, and over 5M sectors transient space available;
but ldev #2 had 1.6M sections transient in use with only 400K available
(and the largest chunk around 256 sectors).
 
By using volutil we changed ldev 2 perm/trans to 100/20 and voila, the
system started allocating transient on ldev 1, and Posix things started
working again!
 
In rethinking this, you probably want to stop allocating transient space
before the volume gets to a fragmented minimum, so perhaps something
along the lines of 70/20 would be better (allowing 10% for small
fragments).  Contiguous space doesn't bother the permanent domain that
much (unless you specify that it should) but it does affect transient
space, particularly Posix-related applications.
 
So the bottom line is... if you WANT the system to make use of ldev 1,
you must restrict the perm/trans allocation of all other system volumes
so that you FORCE it to go to ldev 1.
 
These may be preliminary conclusions... your mileage may vary... etc.
I'll let you know if anything changes.
 
[\] Jeff Kell <[log in to unmask]>