From owner-freebsd-hardware Thu Jun 7 15: 0: 7 2001 Delivered-To: freebsd-hardware@freebsd.org Received: from mail.libam.com (094.dsl108154.surewest.net [63.108.154.94]) by hub.freebsd.org (Postfix) with ESMTP id 793ED37B401 for ; Thu, 7 Jun 2001 14:59:57 -0700 (PDT) (envelope-from MichaelV@EDIFECS.COM) Received: from ecx1.edifecs.com (mail.edifecs.com [207.153.149.131]) by mail.libam.com (8.9.3/8.8.7) with ESMTP id PAA20274; Thu, 7 Jun 2001 15:08:40 -0700 Received: by ecx1.edifecs.com with Internet Mail Service (5.5.2653.19) id ; Thu, 7 Jun 2001 15:04:39 -0700 Message-ID: <36F7B20351634E4FBFFE6C6A216B30D54C26@ecx1.edifecs.com> From: Michael VanLoon To: Michael VanLoon , "'Achim Patzner'" Cc: hardware@freebsd.org Subject: RE: Casing wanted Date: Thu, 7 Jun 2001 15:04:38 -0700 MIME-Version: 1.0 X-Mailer: Internet Mail Service (5.5.2653.19) Content-Type: text/plain; charset="iso-8859-1" Sender: owner-freebsd-hardware@FreeBSD.ORG Precedence: bulk List-ID: List-Archive: (Web Archive) List-Help: (List Instructions) List-Subscribe: List-Unsubscribe: X-Loop: FreeBSD.org OK so I can't do math when in a hurry... heh. 4 x 160 = 640, not 740MB/s... > From: Michael VanLoon > Sent: Thursday, June 07, 2001 1:56 PM > > Proof of concept is fine. I wish you luck on building this > server and its > long-term reliability. So with that, I'm not trying to > convince you not to > build it, but I figure I should address the points I already > brought up. > > > From: Achim Patzner [mailto:ap@bnc.net] > > Sent: Thursday, June 07, 2001 1:24 AM > > > > > Agreed it's quite expensive but... with hardware SCSI > RAID you get: > > > - You don't have all these squirrelly issues you just > > brought up (cable > > lengths, etc.) > > > > Which is about the only problem I encountered and 3ware just > > told me (after > > a direct hit on their long term memory with a hard object) > > that there are > > cables > > up to 70 cm. > > Yes but as far as I know these are outside the ATA100 spec, > and are not > guaranteed to work in all cases. It's kinda hit-and-miss. > If it works, > you're good (unless it starts acting strange). If it doesn't > work, oh well. > Correct me if I'm wrong on this point -- I don't have an > ATA100 spec in > front of me. > > > > - Automatic unattended hardware failover to hot-spare(s) > > > - Automatic unattended background fill-in of data on > > failed-in hot spares > > > while server is live > > > - Caching controller that does delayed elevator-sorted > > write-backs, and > > > read-ahead > > > > This is nothing the 3ware controller won't do. > > Well that hasn't been proven to me. 3ware's specs and > white-papers are > unfortunately alarmingly light on the actual details of how > complete their > live hot-spare support actually is. Although this doesn't > confirm that they > don't do it as well, it's at least noting that they don't > spend any time > saying more than a bullet point "hot-spare". > > A good hardware SCSI RAID controller will fill in the > hot-spare while the > server is up and live. Without the server even realizing > anything happened > (aside from reduced disc performance) it will back-fill the > redundant data > onto the new drive and bring it into the array. Once the new drive is > filled, it is a full member of the array and the machine is > running exactly > as it was before the failure. All this without a reboot, and > without any > downtime. > > One other thing I forgot to mention is dynamic expansion of > arrays. If you > need to add more drives, it will use the same technology to > expand the size > of the array without having to actually move any data (it > will redistribute > the data on the disks for optimal striping, but that is > invisible to the > user or OS). Then if you're running NT you just tell it to expand the > Volume, and if you're running FreeBSD you just run growfs. > Or make a new > partition there if that suits you. > > Without detailed white-papers it's hard to claim the 3ware > controller can do > all this. Maybe it can but I don't see any proof of it in their > documentation. > > > > - Better reliability (yes SCSI drives really ARE built better) > > > > No. I had 15 IBM DDYS (of 35) failing after less than 12 > > months and didn't > > lose > > a single Maxtor. > > Everybody is going to have wildly varying accounts of this > over such a small > sample size. I had 6 of 8 Western Digital IDE drives fail > within 3 days of > buying them. But I realize this is just an anomaly and we > got a bad batch. > > Over the span of several years and hundreds or thousands of > drives, I think > you will find the results generally go in SCSI's favor. > > > > - Higher performance (though yes, IDE performance is pretty good) > > > > Hm. I've seen people getting 95 MB/s through a 3ware RAID. > > Don't forget that > > it got a single channel per disk. > > > > All in all this is a reason why they want this machine - > they want to > > compare > > performance... > > Fine, I'm all for comparing performance. Competition is good! > > However keep in mind that modern SCSI controllers are 160MB/s > per channel. > With four channels that's 740MB/s. And that's just standard > Ultra3 LVD, not > fiber-channel. > > > > - Depending on the controller from 15 to 60 drives per controller > > > > Not really. I'm a strong believer in one channel per disk. > > Well you're looking at it from an IDE point of view. IDE > drives REQUIRE one > drive per controller to get decent performance, because IDE is a much > simpler protocol. And four channels gives you fine fault tolerance, > especially if you're using a technology like RAID-10 (also > called 0+1), > where you spread your mirror drives over multiple busses. > > With tagged-command-queuing, disconnect/reconnect, etc. SCSI > drives can > share a bus without stealing all the bandwidth. A modern > drive can't do any > more than around 40MB/s sustained in optimal conditions > anyway (and that > drops significantly if any seeking is involved). 4 drives > per channel = > 160. 4 channels X 4 drives = 16 drives = 740MB/s. That adds > up to higher > performance on my calculator. Of course these numbers are totally > theoretical, and you will get nowhere near that performance > in the real > world, on a production filesystem, whether it be on SCSI or IDE. > > > > - Higher quality cases, hot-swap cartridges, etc. on the market > > > > Definitely not. The best hot swap cartridge I've ever seen was > > an IDE cartridge. I thought someone mixed Dark Vader and the Cylons > > and turned them into a status display > > Status displays don't equal high-quality hot-swap equipment. > > > > So it's not like you're paying more for nothing. There are > > some very > > > substantial benefits, especially in the reliability/uptime > > > department when a disk fails -- no need to bring the > server down, or > > > even be there when it swaps in a hot spare and starts using it. > > > > Nothing I wouldn't get wit IDE too... > > Once again the documentation is too light on all the details > to confirm > this. It may be true but their documentation doesn't give me > enough details > to confirm it. > > Just a counter-point. :-) > > I'm not trying to dissuade you from building this and > verifying it works. > However if you want a true and fair comparison, you need to > open-mindedly > compare it with all that SCSI RAID has to offer and judge the > plusses and > minuses of each platform. > > If you're really interested in comparing, I'd suggest > starting your SCSI > research here: > http://www.adaptec.com/worldwide/product/prodfulldesc.html?prodkey=ASR-3400S &cat=%2fTechnology%2fRAID%2fRAID+for+Mid-Range+Servers > > There are lots of other good SCSI RAID also, this is just one of several. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hardware" in the body of the message