Date: Wed, 16 Jan 2002 12:44:09 -0800 From: Jason Evans <jasone@canonware.com> To: Bruce Evans <bde@zeta.org.au> Cc: David Wolfskill <david@catwhisker.org>, jhb@FreeBSD.ORG, current@FreeBSD.ORG Subject: Re: -CURRENT as of 14 Jan seems slow Message-ID: <20020116124409.K451@canonware.com> In-Reply-To: <20020116163247.L857-100000@gamplex.bde.org>; from bde@zeta.org.au on Wed, Jan 16, 2002 at 05:04:31PM %2B1100 References: <200201160056.g0G0u3A08394@bunrab.catwhisker.org> <20020116163247.L857-100000@gamplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Jan 16, 2002 at 05:04:31PM +1100, Bruce Evans wrote: > On Tue, 15 Jan 2002, David Wolfskill wrote: > > > >Date: Tue, 15 Jan 2002 16:46:17 -0800 (PST) > > >From: John Baldwin <jhb@FreeBSD.org> > > >Two questions: > > > > >1) Do you have WITNESS on in your kernel config? > > > > Yes, in both the build machine & the laptop -- since before I made a > > "local" hierarchy within my CVS repository (September 9, 2001). > > > > >2) If yes, have you tried building with a kernel without witness? > > > > No, not since I put it in to re-sync with GENERIC. I could try that, I > > suppose -- but as noted, I've had WITNESS in there for a while; something > > seems to have changed during that one 24-hr. period that affected things > > rather radically. And I thought it notable. :-} > > > > I gather no one else has noticed this? > > File locking seems to cause only the usual few percent of slowdown for > each round of major locking changes. I haven't completed benchmarking > the file locking pessimizations. I don't use WITNESS or INVARIANTS > for benchmarking of course. Maybe the file locking changes cause much > larger pessimizations when WITNESS is turned on than most locking > changes. I can see how they might: WITNESS seemed to slow down creation > and destruction of mutexes more than most mutex operations last time I > checked, and there is a descriptor for each file and each file descriptor. Note that additional locking with witness turned on can drastically affect performance. Chances are that Alfred's changes in combination with witness are what caused the slowdown. During certain stages of the lockmgr conversion to mutexes, I saw similar performance degradations (a factor of ~5-10, IIRC). Jason To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20020116124409.K451>