*BSD News Article 59523


Return to BSD News archive

Path: euryale.cc.adfa.oz.au!newshost.anu.edu.au!harbinger.cc.monash.edu.au!news.mira.net.au!Germany.EU.net!howland.reston.ans.net!gatech!newsfeed.internetmci.com!in2.uu.net!olivea!strobe!jerry
From: jerry@strobe.ATC.Olivetti.Com (Jerry Aguirre)
Newsgroups: comp.unix.bsd.freebsd.misc,comp.unix.solaris,news.software.nntp
Subject: Re: INND takethis really long pauses problem (more info)
Date: 20 Jan 1996 05:16:25 GMT
Organization: Olivetti ATC; Cupertino, CA; USA
Lines: 33
Distribution: inet
Message-ID: <4dptr9$o0f@olivea.ATC.Olivetti.Com>
References: <30e8909b.1118570368@newsfeed.kdcol.com> <30ec12f8.10926492@news.kdcol.com> <30f9e36c.10513600@newsfeed.kdcol.com> <slrn4fmpin.l7q.hwr@snert.pilhuhn.de>
NNTP-Posting-Host: strobe.atc.olivetti.com
Xref: euryale.cc.adfa.oz.au comp.unix.bsd.freebsd.misc:12472 comp.unix.solaris:57144 news.software.nntp:19007

In article <slrn4fmpin.l7q.hwr@snert.pilhuhn.de> hwr@pilhuhn.de (Heiko W.Rupp) writes:
>This is a problem with the streaming patch - I never saw it that hard,
>but these pauses ocured from the very beginning. Jerry has posted a
>replacement with some fixes some days ago, so this might help here.
>You might look in news.software.nntp

Just to clarify the situation.  The problem I eventually tracked down
involved innd going into a heavily compute bound state.  The length of
time depended on the size of the article being processed and was most
noticable with the 1 megabyte alt.binaries type.

The key symptoms that marked this condition were that innd was using up
all the free CPU cycles and no other innd work was being accomplished.
The patches I recently posted prevent this from happening.  Note that
streaming is still likely to hog innd and make other functions SLIGHTLY
more sluggish.  With streaming innd can process several history lookups
or small articles in one pass thru the loop.

I also see another problem with different symptoms.  I still sometimes
have long delays for "ctlinnd" commands.  The difference now is that
innd is not consuming abnormal CPU resources.  Monitoring the logs shows
that it continues to service incomming articles from several sites.
This only happens under heavy load.  A second server running the same
code but with only one feed never shows this problem.

Thus it is clear that innd is still cycling thru its channel service
loop.  That should be checking for ctlinnd commands once every pass.
It is unclear to me how the streaming patches could be preventing innd
from getting ctlinnd commands when innd is still checking for them.  I
begin to wonder if some kernal/networking bug is being triggered.
Perhaps the volume of queued data prevents the ctlinnd commands from
being passed?  Is something giving priority to TCP processing over unix
domain?