Date: Sun, 25 Aug 1996 09:15:22 -0500 (CDT) From: Joe Greco <jgreco@brasil.moneng.mei.com> To: michaelv@MindBender.serv.net (Michael L. VanLoon) Cc: jgreco@brasil.moneng.mei.com, freebsd-isp@FreeBSD.ORG, mvanloon@microsoft.com Subject: Re: Anyone using ccd (FreeBSD disk striper) for news Message-ID: <199608251415.JAA29173@brasil.moneng.mei.com> In-Reply-To: <199608241935.MAA05511@MindBender.serv.net> from "Michael L. VanLoon" at Aug 24, 96 12:35:46 pm
next in thread | previous in thread | raw e-mail | index | archive | help
> >At $60 a controller I say stuff the machine with controllers and spread > >your disks out over them! (On a PCI system that means 3 SCSI controllers, > [...] > >Your drives then obviously get spread out among the busses. Note: I stripe > >_across_ busses because I intuitively believe that this may give me better > >response. > > Could you give me an example? > > How does this fit your scheme: two AHC2940UW's (I can probably get > these easier than NCR controllers -- cost isn't a significant factor) > with tagged-command-queuing enabled, four drives (2-4GB), two per > controller. If I put a single ccd across all of them, going in the > order 1, 3, 2, 4. Does that sound like a fairly well optimized start? > > Or, maybe even three AHC2940UW's with six drives, 1, 4, 7, 2, 5, 8, 3, > 6, 9, in a single large ccd. > > Once you star getting multiple ccd filesystems in the news spool, > things become much more complicated (keeping things balanced between > different filesystems). I never liked 'single large' news filesystems because it gets so hard to tell who is filling up your disks in a timely fashion... Some folks have also claimed that updates of fs metadata can still cause various contention problems and associated slowdowns. I've never tried to conduct tests under FreeBSD to see if there is any merit, but even so, I do not like the idea of my entire news spool being GONE if one drive vaporizes on me (I recently lost /news on a machine with separate partitions for alt and alt.binaries at a LARGE ISP, they told me they received few complaints that everything else was gone.. nobody cares about all the good hierarchies anymore). You could go with 3940's (two busses, one card). Unfortunately I noticed some disturbing differences in performance under 2.1.0R between the NCR and the 3940, the NCR was hands down faster... dunno what the current state of affairs is. There's also a 3985 with three busses, and apparently it would not be hard to support the busses but FreeBSD doesn't support it yet. The 3985 also has RAID hardware built in, but driver support would be required, so I would probably consider it to be a three-SCSI-bus controller and nothing more. As for how I lay down filesystems across busses... well it's not too hard, you'll never hit optimum so I just sorta "do it", bearing in mind that swap drives, NOV drives, /usr/local drives, and alt.binaries drives will be the worst bandwidth eaters and should ideally be spread as evenly as possible. I just "start at the top" and assign drives across the busses. For example.. Bus 0 Bus 1 Bus 2 ------ ------ ------ 2G root 2G var news CCD news CCD news/.0 CCD news/.0 CCD usr/local CCD usr/local CCD nov CCD nov CCD alt.binaries CCD alt.binaries CCD A few notes: the stripe size for /usr/local should be *small* because you are really interested in increasing tps to a single large file (history). The alt.binaries stuff should be on a separate CCD because you do not want binaries eating up your nice expensive CCD array of small disks... I usually use two Barra 4G's for binaries. ... JG
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199608251415.JAA29173>