Original problem statement:
orig: Nice works differently under Solaris 2.3 than SunOS 4.1.3 and I would
orig: like to find out if there is any way we can get more control of
orig: processes under Solaris.
orig:
orig: The situation we face is illustrated by the following scenario. We
orig: have two processes, A and B, that are both processor-intensive.
orig: Empirically, under 4.1.3, if A is running and B is niced to 10, A will
orig: get the lion's share of the processor though B will continue to run
orig: with about 15% of the processor. If B is niced to 19, B is effectively
orig: put to sleep until A relinquishes the processor.
orig:
orig: Under Solaris, A and B will share the processor with about a 60%/40%
orig: split regardless of the nice level of B.
orig:
orig: The 4.1.3 nicing is desirable because it allows the user much more
orig: control of each process' access to the processor.
orig:
orig: This is a recurring and significant problem for our shop; if there are
orig: any known patches which address this issue, could you please give me
orig: some pointers about how to improve the situation.
Thanks to:
From: nlf@aluxpo.att.com (Nelson Fernandez)
From: grs@claircom.com (Gregg Siegfried)
From: itubjs@lux.levels.unisa.edu.au (Bevin Steer)
From: stern@sunrise.East.Sun.COM (Hal Stern - NE Area Systems Engineer)
From: Eckhard.Rueggeberg@ts.go.dlr.de (Eckhard Rueggeberg)
Short story:
Wait for Solaris 2.4.
Long story:
When the prioritization works correctly, it will give very nice control
and flexibility for manipulating jobs and job queues. Hal, of course,
sent a complete and marvellous explanatory article on the power of
Solaris job control (SunWorld, July 1993). Briefly, it relies on two
easily accessible utilities: priocntl(1) and dispadmin(1m). Further,
S/A's have the option of creating local job classes for fine tuning
process control.
We tried priocntl on our target applications, though, and behaviour was
identical to the problems we have been having with the current version
of nice (nice values change, priorities do not). Eckhard pointed out
that Sun does indeed recognize that this is a bug and has scheduled the
fix to appear in the 2.4 release.
sid
S. Cowles
voice: +1 415 812 1917
fax: +1 415 812 1916
email: scowles@shgc.stanford.edu
email: scowles@llnl.gov
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:09:11 CDT