Date: Mon, 17 Sep 2001 08:20:12 -0700 (PDT) From: Matthew Jacob <mjacob@feral.com> To: Terry Lambert <tlambert2@mindspring.com> Cc: Christoph Hellwig <hch@ns.caldera.de>, Dennis Berger <Dennis.Berger@nipsi.de>, <freebsd-fs@FreeBSD.ORG>, <opengfs-devel@lists.sourceforge.net> Subject: Re: Porting a new filesystem to FreeBSD Message-ID: <20010917080710.R58734-100000@wonky.feral.com> In-Reply-To: <3BA5E3E2.AB247A2C@mindspring.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 17 Sep 2001, Terry Lambert wrote: > Christoph Hellwig wrote: > > Nice thing but irrelevant until SCSI III is actually in more > > widespread use. > > Seagate supports it today. I'm 65% certain IBM does as well; > the issue with SCSI III is that they haven't been able to make > the optical parts fast enough (same as the issue with 10Gbit > ethernet, which you can buy from HP today). A slight clarification here. SCSI-3 is a set of specifications about various forms of the SCSI protocol. This has been pretty much put to bed some time back. Various pieces of it evolve. The commands for devices are in various subportions- DLOCK would probably be described in SPC-2 (SCSI Primary Commnds, 2) which is ahead of the nominal 'SCSI-3' set. The actual hardware transport for old fashioned copper wire SCSI is more formaly known as SPI (SCSI Parallel Interface). This is under the T10 part of NCITS (http://www.t10.org). Fibre Channel, at the signalling level, is covered under FC-PH and FC-PH-2, and is covered under the T11 part of NCITS (http://www.t11.org). The specs for each develop pretty nearly independently. The current 'just now shipping' parts for optical FC transport is 2Gb. Considering that *most* system I/O bus interfaces will practially saturate with this, worrying about 10Gb yet is premature. At any rate, locking, which this is the primary issue of, is not sensitive to data rate. It's more sensitive to media reliability and packet overhead. SCSI over FC vs. IP over FC or IP over GigEthernet has higher overhead (due to the nature of FC-SCSI exchanges vs. IP Exchanges), SCSI over FC has substantially less overhead than SCSI over SPI. One of things I've argued for is that DLOCK commands should piggy back with I/O commands. That is, you do a write of metadata along with the lock name and acquisition info- if you acquire the lock, the write succeeds- basically like the alpha's Store Locked Conditional instructions. Or you could use a single FC ELS frame to do the DLOCKs and not even have it part of SCSI at all. Or you could use a parallel port scoreboard interconnect (somebody made a cluster of Multias using this). There are a variety of locking mechanisms. > > > > I might as well use XFS, then, which is at least being ported to > > > FreeBSD... > > > > But XFS is not distributed filesystem. (So if the port is actually There's cXFS... -matt To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-fs" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20010917080710.R58734-100000>