Date: Fri, 14 Oct 2005 18:34:51 +0200 From: Andre Oppermann <andre@freebsd.org> To: Eric Anderson <anderson@centtech.com> Cc: freebsd-current@freebsd.org Subject: Re: ufsstat - testers / feedback wanted! Message-ID: <434FDE2B.8090502@freebsd.org> In-Reply-To: <434FDAB2.7040402@centtech.com> References: <200510131412.23525.max@love2party.net> <20051013181026.GB27418@odin.ac.hmc.edu> <20051014091004.GC18513@uk.tiscali.com> <20051014.085816.104604949.imp@bsdimp.com> <434FDAB2.7040402@centtech.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Eric Anderson wrote: > M. Warner Losh wrote: > >> In message: <20051014091004.GC18513@uk.tiscali.com> >> Brian Candler <B.Candler@pobox.com> writes: >> : On Thu, Oct 13, 2005 at 11:10:26AM -0700, Brooks Davis wrote: >> : > > I don't think you can measure one single interger (or 64bit) >> increase in face : > > of a operation that has to access backing >> store. Even if there is a : > > performance hit, you don't have to >> build your kernel with the option enabled. >> : > : > The one thing I'd be worried about here is that 64bit updates are >> : > expensive on 32bit machines if you want them to be atomic. >> Relative to >> : > backing store they probably still don't matter, but the might be >> : > noticable. >> : : I'd be grateful if you could clarify that point for me. Are you >> saying that >> : if I write >> : : long long foo; >> : ... >> : foo++; >> : : then the C compiler generates code for 'foo++' which is not >> thread-safe? >> : (And therefore I would have to protect it with a mutex or critical >> section) >> : : Or are you saying that the C compiler inserts its own code around >> foo++ to >> : turn it into a critical section, and therefore runs less efficiently >> than >> : you'd expect? >> >> You have to protect this thread-unsafe operation yourself. > > > For statistics gathering purposes though, should I worry about this, or > go for 'fast and imperfect' instead of 'perfect and slow'? With > filesystems, I think it's more important to leave performance high and > get a notion of the statistics, rather than impact performance for > perfect stats (that you may only look at occasionally anyhow). If you make it a #define macro then you can leave the choice for the compile time. Fast and lose when i++ and safe and slow when atomic_inc(&i, 1). -- Andre
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?434FDE2B.8090502>