Date: Tue, 3 Jun 2003 17:32:04 +0100 From: Paul Robinson <paul@iconoplex.co.uk> To: Matthew Dillon <dillon@apollo.backplane.com> Cc: arch@freebsd.org Subject: Re: Making a dynamically-linked root Message-ID: <20030603163204.GA29331@iconoplex.co.uk> In-Reply-To: <200306031614.h53GEqkU008308@apollo.backplane.com> References: <20030602171942.GA87863@roark.gnf.org> <xzpznl02nry.fsf@flood.ping.uio.no> <20030603080456.GA57773@cirb503493.alcatel.com.au> <200306031614.h53GEqkU008308@apollo.backplane.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Jun 03, 2003 at 09:14:52AM -0700, Matthew Dillon wrote: > I'm fairly sure there isn't an issue. Both a hard drive's own > on-board cache and FreeBSD's clustering and caching code are *very* well > suited to this sort of parallel initiation. There is certainly no > scheduler issue. The key advantage here is that you are removing > serialization that would otherwise cause both cpu cycles and disk > cycles to be wasted waiting for each other. Take sendmail for example. > sendmail usually takes upwards of a second to startup due to initial > DNS lookups that it makes and other things. sshd doesn't start > instantaniously either, I think due to creating the initial > session keys. I'm sorry, I think I have to point something out here: We're talking about shaving a few seconds off a process that on a well-maintained server happens once a week at most - i.e. a cvsupdate, buildworld, install-kernel, reboot on a Monday morning. That's a few seconds a week. This is a lot of effort for shaving a few seconds off a week. It's all well and good talking about getting rid of "wasted cpu cycles and disk cycles" for something that happens perhaps a few times a second, but once a week? And this is to give us a performance advantage over Linux? Not being funny, but is this another troll thread? If you really want to speed up boot times because you're moving a laptop around a lot, great, there is a possibility that an argument can be made that someone could be looking at the serialisation of tasks that assist in bootstrapping other processes - e.g. a dns resolver process that gets all the "usual" hostnames into cache, key initiation stuff, etc... or of course, we could look at getting rid of the incredibly serial rc scripts for certain tasks so lot's of things bootstrap at once - providing dependancies are tracked, and then we get into a discussion akin to package management. Interesting discussion, but are there not other areas we could look at for performance gains? Or am I just not "getting it"? -- Paul Robinson
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030603163204.GA29331>