Date: Sat, 04 Dec 1999 15:02:18 -0800 From: Kirk McKusick <mckusick@flamingo.McKusick.COM> To: George Michaelson <ggm@dstc.edu.au> Cc: Mike Smith <msmith@freebsd.org>, match@elen.utah.edu, current@freebsd.org Subject: Re: Mounting one FS on more than one system Message-ID: <199912042302.PAA05208@flamingo.McKusick.COM> In-Reply-To: Your message of "Sun, 05 Dec 1999 09:44:32 %2B1000." <1082.944351072@dstc.edu.au>
next in thread | previous in thread | raw e-mail | index | archive | help
To: Kirk McKusick <mckusick@flamingo.McKusick.COM> cc: Mike Smith <msmith@FreeBSD.ORG>, match@elen.utah.edu, current@FreeBSD.ORG Subject: Re: Mounting one FS on more than one system In-reply-to: Your message of "Sat, 04 Dec 1999 12:44:43 PST." <199912042044.MAA05073@flamingo.McKusick.COM> Date: Sun, 05 Dec 1999 09:44:32 +1000 From: George Michaelson <ggm@dstc.edu.au> Let me start by saying that I do not have the definitive answers to all your questions. I'll take a crack at some of them. Allowing for cache writeback delays, is the speed of direct-to-shared-disk fast enough that using NFS as an "abstraction" layer would be faster than any network extant? The gains come from being able to read data directly from the disk rather than transferring it across the network. However, the cost of maintaining cache coherency would be at least as difficult and bandwidth consuming as a distributed filesystem. Would it be as fast? would the effort to make this work exceed the cost of making real networks exist? As the network speed approaches the disk speed, the gains would diminish. It would seem that there might be opportunities to do 'cut through' in the coding for known-private files after open (ok, inode allocation/extension has problems) to optimize them to at-worst 'disk+bits' instead of NFS costs. The problem is in identifying when private goes to shared. Also as you point out, new block and inode allocations have to be centrally coordinated. If one party mounts -r the FS (eg news spool) then does this reduce the complexity? eg /usr mounted read-mostly for a bunch of tightly coupled boxes. If any machine can write, then all the other machines have to have some way of keeping their caches consistent with the machine that did the modification. If some other protocol is used for interlock, does this make mmap shares across clusters faster? Mmap sharing across machines is going to be slow. I have never been a fan of distributed shared memory as a programming model, and this does not look like a way of making it run any faster. -George -- George Michaelson | DSTC Pty Ltd Email: ggm@dstc.edu.au | University of Qld 4072 Phone: +61 7 3365 4310 | Australia Fax: +61 7 3365 4311 | http://www.dstc.edu.au 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?199912042302.PAA05208>