I received more responses to this than I expected, and I thank each and
every one of you who replied.
Worthy of note is that advfsd seems to be well known for causing
performance problems, and eating cpu cycles. We're not using that daemon
on our system, so it isn't an issue in our case, but certainly worth
noting.
Apparently joind has also been known to runaway. Again not an issue in
our case, but it was recommended to be replaced with the old bootpd, if
needed.
Many replies suggested various pieces of software (listed below with
sites to download from) which might be useful to try and find out more
information about where my cpu cycles are going. I will have a look at
these, hoping that even one can point something out I'm not seeing.
Ole Craig (<olc_at_cs.umass.edu>) recommended that we consider using Dnews
rather than Inn for our news server. He has seen performance
improvements at his site since switching to Dnews. This will be worth
looking into, I'm sure.
John H. Warren (<JHWARREN_at_ESCOCORP.com>) suggests that DU-4.0d will have
configuration options which will make it easier to configure for
specific purposes, including news service.
Dave Tetreault (<davet_at_uriacc.uri.edu>) confirms that Inn and AdvFS
don't really get along very well. There are common reports about this,
but I also got a response which indicates that it *can* work well:
Richard L Jackson Jr (<rjackson_at_portal.gmu.edu>) has a rather nice
system compared to our humble news server:
RJ> We have a 2100 4/275 with three CPUs, 2GB memory, and eight
RJ> DS-RZ40-VA (9GB) disk drives attached to a SWXCR (EISA) configured
RJ> as two RAID 0+1 logical drives running Digital UNIX 4.0B for our
RJ> USENET, Listproc, primary DNS, NTP, bootp, whois services and find
RJ> the system happy. Of course, there is little to no free memory but
RJ> the CPUs are well balanced (We use AdvFS 100%) with a low number of
RJ> memory page outs.
In further communication, Mr. Jackson comments:
RJ> I sized the 18GB news spool volume and all our volumes for the
RJ> number of files I think we will have using mkfdmn -p flag to
RJ> preallocate the metadata. So far it has worked well.
He also suggests switching from immediate swap to deferred swap, by
moving (or deleting or renaming) the file (actually a sym-link to the
primary swap partition) /sbin/swapdefault which is used to set the
system for immediate swapping. Details are in the swapon(8) manpage.
He also asked me for more information, which I forgot to send. :-(
Kurt Carlson (<snkac_at_java.sois.alaska.edu>) recommended several of the
programs listed below, and also provided a ksh script which has provided
me with insight into using dbx on the kernel to gather performance data.
Unless I'm misunderstanding, this script (vmstat.ksh) is available in
the uaio distribution, (see below).
Anthony Talltree (<aad_at_nwnet.net>) recommends keeping control of the
number of articles in some of the busier newsgroups to keep directory
access times down. He also suggested striping the news spools, to
improve disk performance.
AT> Don't mess with RAID5 -- just use a plain nonredundant stripe.
AT> Turn off access-time updates on the spool filesystem if you can.
I thank each and every one of you for your responses. Essentially it
comes down to needing to find out more about what the system is doing,
and I'll be looking at that for the next little while.
----------------------------------------------------------------------
Sylvain Robitaille syl_at_alcor.concordia.ca
Assistant to the System Managers Concordia University
Computing Services Department Montreal, Quebec, Canada
----------------------------------------------------------------------
Programs which have been recommended:
Monitor: use instead of top. Expands system time by showing the
disk/io wait time.
Available on free source code cd's from digital or on their
ftp server.
Dnews: News server software.
www.netwinsite.com
uaio: rewritten iostat to include all disks and better summary
(including by bus):
ftp://raven.alaska.edu/pub/sois/README.uaio
ftp://raven.alaska.edu/pub/sois/uaio-v1.8a.tar.Z
uakpacct: much like acctcom but with more filters (have daily
email reports of threshhold users sent automatically):
ftp://raven.alaska.edu/pub/sois/README.uakpacct
ftp://raven.alaska.edu/pub/sois/uakpacct-v1.6.tar.Z
uaklogin: combines attributes of ps, finger, and pstree under
one command to get a view by terminal by process tree.
same kit as uakpacct.
proc_info: for getting details on a specific process including
full resource utilization and env variables.
ftp://raven.alaska.edu/pub/randy/proc_info_utils/
vm_mon: for ubc and advfs monitoring.
ftp://raven.alaska.edu/pub/randy/perf_mon_tools/vm_mon-1.0.README
ftp://raven.alaska.edu/pub/randy/perf_mon_tools/vm_mon-1.0.tar.gz
Kurt Carlson (<snkac_at_java.sois.alaska.edu>) warns regarding vm_mon:
KC> I'm not sure Randy has this in full production state. The problem
KC> is advfs bypasses UBC and vmubc misses that and doesn't properly
KC> report anyway. The vmstat.ksh example script in uaio distribution
KC> includes a dbx logging call which also grabs ubc info (but not the
KC> advfs buffers which are separate from ubc).
KC> under v4.0 there is also an advfsstat, but you have to know which
KC> domain you want for that to be of any use.
Received on Thu Jan 08 1998 - 00:49:52 NZDT