From owner-freebsd-stable Fri Feb 2 12:03:32 1996 Return-Path: owner-stable Received: (from root@localhost) by freefall.freebsd.org (8.7.3/8.7.3) id MAA19185 for stable-outgoing; Fri, 2 Feb 1996 12:03:32 -0800 (PST) Received: from dtr.com ([204.119.17.55]) by freefall.freebsd.org (8.7.3/8.7.3) with SMTP id MAA19158 for ; Fri, 2 Feb 1996 12:02:09 -0800 (PST) From: bmk@dtr.com Received: (from bmk@localhost) by dtr.com (8.6.12/8.6.9) id LAA14322 for stable@freebsd.org; Fri, 2 Feb 1996 11:14:26 -0800 Message-Id: <199602021914.LAA14322@dtr.com> Subject: Re: New version of ccd driver available To: stable@freebsd.org Date: Fri, 2 Feb 1996 11:14:25 -0800 (PST) In-Reply-To: <199601311134.DAA07361@silvia.HIP.Berkeley.EDU> from "Satoshi Asami" at Jan 31, 96 03:34:52 am Reply-To: bmk@dtr.com X-Mailer: ELM [version 2.4 PL24] MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: owner-stable@freebsd.org Precedence: bulk > (7) What's the second field in /etc/ccd.conf? > That's the "interleave size". Basically, the ccd driver will write > this many sectors (usually 512 bytes) to one disk before it moves to > the next disk. As a special case, a zero here means no interleave, > i.e., to concatenate disks serially. > We have found that in FFS, a value of 16 usually optimizes read > performance, while the write peaks with a much larger value (like > 512). This probably has to do with cluster_write() thinking it's > writing to a single disk when it's actually not. This is one of the > things we are planning to fix. I've done some performance testing - both mirrored and striped - on one of my systems. I noted a different performance curve than you did, so I thought you might like to see them. (I used a SCSI-I disk, and provided the results run on a standard FFS - the numbers are mostly useful to compare to the FFS baseline results.) The results are at "http://www.dtr.com/ccd" - Mostly it's raw results and a quick-n-dirty summary and Excel spreadsheet with the performance curve graphed. The results and summary are in ASCII. I have also included the script that I used to run the tests. I've also got some ambitious ideas for improving the ccd driver - nothing coded, just some raw ideas. (hot swappable mirrors, etc.). At work, I deal with three different host-based fault-tolerance implementations (Sequent Dynix and ptx/SVM, as well as the Solaris disk suite) - the current ccd driver resembles Dynix; some of the ideas I have in mind are borrowed from SVM and Solaris. If you're interested in hearing what I have in mind, I'll write something up and forward it to you.