Date: Tue, 10 Dec 2002 17:08:13 +1030 From: Greg 'groggy' Lehey <grog@FreeBSD.org> To: Mark Rowlands <fuc952d@tninet.se> Cc: questions@FreeBSD.org Subject: Re: couple of vinum questions from a vinum newbie. Message-ID: <20021210063813.GC94148@wantadilla.lemis.com> In-Reply-To: <200212100708.15890.fuc952d@tninet.se> References: <200212081704.28492.fuc952d@tninet.se> <20021208234555.GO96646@wantadilla.lemis.com> <200212100708.15890.fuc952d@tninet.se>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tuesday, 10 December 2002 at 7:08:15 +0100, Mark Rowlands wrote: > On Monday 09 December 2002 12:45 am, Greg 'groggy' Lehey wrote: >> On Sunday, 8 December 2002 at 17:04:28 +0100, Mark Rowlands wrote: >>> I have decided to dip my toe into the vinum waters primarily for improved >>> read performance, not redundancy or reliability. I have backups for that >>> ;-). >>> >>> I have two questions, one related to optimal stripe size and one to >>> gathering statistics. >>> >>> I have a freshly cvsupped 4.7 stable currently installed on a 30g single >>> disk I propose to add a second 30gb disk as a mirror of this. and to add >>> two new identical disks (maxtor fdb-9 120gb) and create a striped volume >>> for data on those. >>> >>> I propose (maybe) to create a single striped volume but...... what stripe >>> size? Reading the various vinum docs and looking at some of the samples >>> I found lying around the web left me a touch confused. >>> >>> There are two categories of files on the data disk, largeish ( 4-5mb) >>> which are most frequently accessed but the overwhelming majority of files >>> are much smaller (<4k) but not accessed so frequently. >>> >>> So, is it better to try and optimize for the most frequently accessed, or >>> the filesize that comprises the bulk of file accesses? or better to >>> create two volumes with different stripe sizes. In either case, if anyone >>> has any concrete recommendations, I would welcome them. >>> >>> When the new hardware arrives, (about a week), I will probably run some >>> tests with varying stripe-sizes and see what happens but any pointers >>> towards likely good starting points...... >> >> There's a fair discussion of stripe sizes in vinum(4), about 200 lines >> long, under the heading "Performance considerations". I can't do >> better than that in less space. What part of that don't you >> understand? >> > yes, that was a poorly phrased question..... more of a bleat than a question > really. What confuses me in particular, is that despite vinum (8) saying that > > "with modern disk sizes and correct raid implementations there is no reason > why stripe sizes of between 2-4mb could not be used" > > and > > "a good value for stripe size is betwween 256 and 512 but avoid powers of > 2..." > > The samples then shown in vinum (8) all then use stripe sizes of 512k (a > power of 2) or 32k or 64k (distinctly less than 256). *sigh* I don't see any stripe sizes of 512k, though there are some 512b (256k), and yes, there are some tiny stripes too. What can I say? I've known for years that powers of 2 are Wrong, but I still haven't got round to updating the docco. I'll do it soon if I remember; I'll remember if somebody enters a PR against it. The truth is that you should choose a stripe size which is large and not a power of 2, though it should be a multiple of the block size of the file system you put on it. Since that's generally 16 kB, but could be 32 kB, something like 480 kB (5123 kB - 32 kB) is about right. Why not have even larger stripes? Superstition, I suppose. If you make them too big, they become concatenated. The real issue is whether it makes any difference beyond 512 kB. > This is also common in the cases where I have found sample > configurations on the internet. > > So I am, as we speak, running some tests with various stripe sizes and newfs > options to help me see the performance differences for myself. Apologies for > wasting your time, I have posted a note on my monitor....."think before you > hit the send button" I will also have a look at ccd and atacontrol. I'd be very interested to see the results of rawio on the various configurations. Greg -- When replying to this message, please copy the original recipients. If you don't, I may ignore the reply or reply to the original recipients. For more information, see http://www.lemis.com/questions.html See complete headers for address and phone numbers To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-questions" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20021210063813.GC94148>