From owner-freebsd-hackers Thu Apr 6 22:16:55 1995 Return-Path: hackers-owner Received: (from majordom@localhost) by freefall.cdrom.com (8.6.10/8.6.6) id WAA22965 for hackers-outgoing; Thu, 6 Apr 1995 22:16:55 -0700 Received: from gndrsh.aac.dev.com (gndrsh.aac.dev.com [198.145.92.241]) by freefall.cdrom.com (8.6.10/8.6.6) with ESMTP id WAA22958 for ; Thu, 6 Apr 1995 22:16:50 -0700 Received: (from rgrimes@localhost) by gndrsh.aac.dev.com (8.6.8/8.6.6) id WAA06264; Thu, 6 Apr 1995 22:16:32 -0700 From: "Rodney W. Grimes" Message-Id: <199504070516.WAA06264@gndrsh.aac.dev.com> Subject: Re: large filesystems/multiple disks [RAID] To: smpatel@wam.umd.edu (Sujal Patel) Date: Thu, 6 Apr 1995 22:16:31 -0700 (PDT) Cc: hackers@FreeBSD.org In-Reply-To: from "Sujal Patel" at Apr 7, 95 00:30:09 am X-Mailer: ELM [version 2.4 PL23] Content-Type: text Content-Length: 2497 Sender: hackers-owner@FreeBSD.org Precedence: bulk > > > On Thu, 6 Apr 1995, Rodney W. Grimes wrote: > > > The drive has a 5 year warranty, so it's not like your buying a very > > used drive. Or make me an offer if you think $800 is too much for > > it. > > > > My DAT drive should be in next week so I am not worried about disk > > failure causing me to lose data with stripped disks. > > > > I have read the code in stripe.tar, it should be a day or twos work > > to get it up and running under FreeBSD. > > A good question is, can you get this to improve raw throughput much? I > began implementing striped disks on L*nux a while back with two 730 Meg > Quantum Lightnings on a U24F, and got no improvement in speed at all. I > did implement it at the block device level with a few lower level hacks, > so that may have been a problem with the way I designed it, but what kind > of SCSI card do you expect to push more than 3-5 Megs/Sec. My 24F can't > put out more than 2.3 megs/sec sustained from the FileSystem (I think my > drives should be able to do better than this). Did you have spindle sync on the Quantum drives? This is very important if you want to eliminate rotational delay problems when muxing the data out to the drive. You also need to have the interleave data chunk match closely to the drive cache size (or smaller, but not to small). Your choice of controllers was poor. I have run 2 drives on a 1742 doing iozone to both drives at the same time (3+MB/sec each drive) for a controller throughput of 6MB/sec. I have run 2 4MB/sec drives on an NCR 825 based controller and got 7+MB/sec combined throughtput. Ultrastore controllers have become famous for there in ability to do a decent job of getting data to and from the disk. I suspect that the NCR825 with wide drives could easily hit 15MB/sec, and who knows, maybe even get close to the 20MB/sec wide scsi speed limit. (Any one out there with a pile of fast drives on a fast wide scsi controller that can do some aggregate bandwidth tests??). If the controller gets in my way, I will add controllers, NCR810 cards are cheap at < $85.00 a pop. Even fast+wide 825 controllers are not that bad at <$189.00 each. Now that we have CDDI network cards that can push 8MB/sec I want to bring my disk bandwidth up to a number higher than my network. Can you say diskless PC's that actually *perform*!! -- Rod Grimes rgrimes@gndrsh.aac.dev.com Accurate Automation Company Custom computers for FreeBSD