Date: Fri, 7 Jul 2000 08:43:22 -0700 (PDT) From: Matthew Dillon <dillon@apollo.backplane.com> To: Bill Fumerola <billf@chimesnet.com> Cc: Alfred Perlstein <bright@wintelcom.net>, Marius Bendiksen <mbendiks@eunet.no>, freebsd-arch@FreeBSD.ORG Subject: Re: Alterations to vops Message-ID: <200007071543.IAA00597@apollo.backplane.com> References: <Pine.BSF.4.05.10007062327020.68909-100000@login-1.eunet.no> <200007070143.SAA96248@apollo.backplane.com> <20000706184859.C25571@fw.wintelcom.net> <20000706215543.Z4034@jade.chc-chimes.com>
next in thread | previous in thread | raw e-mail | index | archive | help
:
:On Thu, Jul 06, 2000 at 06:48:59PM -0700, Alfred Perlstein wrote:
:
:> > What may be happening here is stalling in namei(). find and cvsup
:> > are very heavy on path lookups and that combined with seek latency
:> > on the drive could result in filesystem locks on directories being
:> > held for much longer periods of time then normal. Any other process
:> > trying to 'open' a file (verses reading or writing an already-open file)
:> > would start to stall.
:>
:> just a note:
:> sysctl -w vfs.vmiodirenable=1
:> helps a _lot_.
:
:I'll turn it on and see if that helps.
:
:--
:Bill Fumerola - Network Architect / Computer Horizons Corp - CHIMES
:e-mail: billf@chimesnet.com / billf@FreeBSD.org
Ahhhhh. We're waiting with baited breath!
If setting vmiodirenable helps then what's probably happening is
that the buffer cache's malloc space is being blown away by all
the path lookups (due to the reading of all the directories). Turning
on vmiodirenable would greatly increase the amount of memory available
for caching directories.
For the other person who asked what vmiodirenable is: It's a mechanism
to allow the system's VM page cache (the master page cache in the system)
to cache file pages associated with directories. With this turned on
the system is able to cache at least 10x as many directories.
With the option turned off (the default, see note below), the buffer
cache implements a special 'malloc space' to hold directory pages.
This space is very small, and we can't make it any bigger without
really screwing up the system.
sysctl -a | fgrep malloc
My intention was to turn on vmiodirenable by default, but I got stalled
on an obscure bug or two related to softupdates. That was a long time
ago.. at least 6 months, probably more. It could very well be that
the problems no longer exist. My intention is also at some point to
remove the crufty malloc space feature in the buffer cache. It was
originally designed to 'save memory' by packing small filesystem blocks
rather then using the mandatory minimum of one page (4K on i386) that
anything backed by the VM Page cache requires. For example, most
small directories fit in much less then 4K. The concept is sound,
but the implementation has severe scaling problems.
-Matt
Matthew Dillon
<dillon@backplane.com>
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-arch" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200007071543.IAA00597>