From owner-freebsd-hackers Sun Jan 21 15:08:46 1996 Return-Path: owner-hackers Received: (from root@localhost) by freefall.freebsd.org (8.7.3/8.7.3) id PAA12356 for hackers-outgoing; Sun, 21 Jan 1996 15:08:46 -0800 (PST) Received: from localhost.cdrom.com (localhost.cdrom.com [127.0.0.1]) by freefall.freebsd.org (8.7.3/8.7.3) with SMTP id PAA12349 Sun, 21 Jan 1996 15:08:43 -0800 (PST) Message-Id: <199601212308.PAA12349@freefall.freebsd.org> X-Authentication-Warning: freefall.freebsd.org: Host localhost.cdrom.com [127.0.0.1] didn't use HELO protocol To: Jaye Mathisen cc: hackers@freebsd.org Subject: Re: some RAID numbers. In-reply-to: Your message of "Sun, 21 Jan 1996 14:41:25 PST." Date: Sun, 21 Jan 1996 15:08:43 -0800 From: "Justin T. Gibbs" Sender: owner-hackers@freebsd.org Precedence: bulk >On Sun, 21 Jan 1996, Justin T. Gibbs wrote: > >> > >> >Adjile RAID box, 5 Fuji 4GB disks, RAID 5, verify off (on cache), 8MB cache >. >> > >> >P5120, 64MB RAM, FBSD 2.1 stable (supped 1/21), Adapter 2940UW, tag enabled >. >> >> How many tags? The default is only two unless you play around with the >> stuff around QUEUE_FULL_SUPPORTED in i386/scsi/aic7xxx.c > >Whatever is default, I didn't touch anything other than the kernel config >file. Try bumping it to 8. It would be interresting to see what impact it has on your performance. ( make the appropriate line +=6 instead of +=2 inside the QUEUE_FULL_SUPPORTED stuff). >> >> >> These numbers seem rather low. What is the performance of the individual >> fujitsu drives. Running bonnie on a single Quantum Atlas will blow >> these numbers away. Are you doing RAID 5? Even with RAID 5, I would >> still expect much better numbers. > >Well, I ran raid 0, and it hauled butt. These numbers compare pretty >well with a RAID box we have on an alpha under OSF with a mylex card. >I've never seen a RAID 5 that can do better than single spindle speed, it >would seem difficult to do, given the read write cycle. I take it you have spindle sync enabled and have played with the rotational offsets for the drives (how far behind the master a drive is)? Rod would have more information on how to do this optimally(rgrimes@FreeBSD.org). I would expect for a RAID 5 box to give much higher performance than a single drive. Calculation of parity can be overlapped with the SCSI controller's transfer of data to the RAID box. Of course you do lose 1/5 of your bandwidth, but even so, reads should be as fast as a 4 disk RAID 0 box (no parity calculartions unless there is an error retriving data from a drive) and writes should only be some fraction slower than a 4 disk RAID 0 setup. >Note also that I did say the FS was partially used, nor was the box >completely idle. (although there were no processes accessing that >controller/card other than mine. > >I'd like to see the output from bonnie, just so I have some comparisons. >I have a barracuda with a 2940, and I don't see much better numbers. I don't have an Atlas, but there are many people on this list that do. Hopefully one of them will give more exact numbers. I don't recall the Bonnie results, but iozone for a 2940W to a wide Atlas was ~7.2MB/s with a command overhead of ~430us (using Bruce's disklatency program). -- Justin T. Gibbs =========================================== FreeBSD: Turning PCs into workstations ===========================================