Re: Mirroring

}From: ckipp@vth1.vth.colostate.edu (Cathy Kipp) }Subject: Re: Mirroring }To: johnl@informix.com }Date: Thu, 18 Aug 1994 09:34:19 +22305606 (MDT) } }Hi John, } }Given all of the interest this topic seems to be getting, would you mind }if I sent out your reposnse to me publicly? Or perhaps you would like to }do it and then you could edit [it]. It just seems like a lot of us have }been listening to the marketing types, and I think a lot of people missed }what you were trying to convey in your initial message, which you }subsequently explained so well to me. Being easily swayed by flattery, I'm sending a mildly edited version of the response Cathy refers to to comp.databases.informix. The subject of LVMs and performance is complex and controversial (and important). I have a file of some 2800 lines at the moment recording various discussions on the subject -- and that's after I've weeded out a fair number of gratuitous repeats and 10-line signatures, and so on. I also attach John Bria's note which provides useful ammunition for my contention. One of the claimed advantages for DSA disk handling is that it can selectively read from either the primary or the secondary of a mirrored pair -- so can the O/S mirrored systems, and here is evidence that at least one does precisely that. Yours, Jonathan Leffler (johnl@informix.com) #include <disclaimer.h> }From johnl Wed Aug 17 12:08:31 1994 }To: ckipp@vth1.vth.colostate.edu }Subject: Re: HELP - Hardware mirroring versus dbms mirroring } }Hi Cathy, } }Yes, I've heard the same story. I simply haven't yet heard a convincing }explanation of why our software can do it better than the O/S can. } }It does, of course, depend on the mirroring support from the O/S, and }especially on a multi-CPU machine where the O/S I/O's are handled by }a single CPU under the native mirroring system, DSA could have an }advantage if it has multiple threads handling the writes in parallel. }But I'm not convinced that O/S mirroring is that bad. It also depends }on the intelligence or otherwise of the disk controllers. Etc. } }I stand by my observation [...] that unless the O/S has screwed up badly, I }don't think that the Informix mirroring provides much (if any) advantage. } }I don't have any concrete evidence either way, and it is very difficult to }determine experimentally. I know that there were once plans at one time to }spend a day or two assessing the effect of LVMs (logical volume managers) }on the performance of OnLine. I also know that it didn't happen -- I hope }it was in part because I pointed out to the person who was asked to do the }test that controlling the parameters of the test was going to be difficult, }and was going to need considerably more than a day or two simply to work }out what to test and how, independently of the time taken to create and }load suitable data sets (mainly large ones) under multiple different }configurations with differing amounts of RAID-ness, different numbers of }controllers, different places where the mirroring occurs, different numbers }of CPUs, different numbers of AIO threads, etc. I should have mentioned striping as another factor which adds to the complexity... }So, yes, I think maybe you are being lead astray by listening to Informix }marketing talk. Summary of previous discussion... }>From: ckipp@vth1.vth.colostate.edu (Cathy Kipp) }>Subject: Re: HELP - Hardware mirroring versus dbms mirroring }>To: johnl@informix.com }>Date: Wed, 17 Aug 1994 12:50:23 +22305606 (MDT) }> }>Jonathan Leffler writes: }>->}From: ckipp@vth1.vth.colostate.edu (Cathy Kipp) }>->}Subject: Re: HELP - Hardware mirroring versus dbms mirroring }>->}Date: Wed, 17 Aug 1994 09:03:53 +22305606 (MDT) }>->}X-Informix-List-Id: <list.4531> }>->} }>->}Joe Lumbley writes: }>->}->Hardware mirroring is usually best, as it's faster. Next is HP-UX. Last }>->}->comes OnLine mirroring. OnLine mirroring contains a little bit of logic }>->}->regarding how to handle chunks that are down, but with any luck the HPUX }>->}->or hardware mirroring will never let OnLine see that situation anyway. }>->} }>->}Joe, }>->}Not to disagree with an expert, but I think it's important to mention here that }>->}while hardware mirroring may be faster with the OnLine engine, under the 6.0 }>->}DSA engine, Informix miroring will offer some performance advantages. }>-> }>->I haven't been controversial for a while, so why not now... }>-> }>->Jon Bentley (or someone like him) said somewhere words to the effect that }>->'if you can simulate something in the language faster than the compiler }>->does the same operation, the compiler writer has blown it badly'. }>-> }>->I think much the same applies to disk mirroring -- if the O/S can't do it }>->better than OnLine, then the O/S mirroring is probably defective. }> }>Okay.... I suppose I should have said: }>"Informix mirroring under DSA is SUPPOSED to offer some performance advantages". }> }>Or so I've heard. Perhaps it's not really the case, but I have heard that }>under DSA Informix, when using mirroring, Informix will read from the }>mirrored disk, which is supposed to help the speed of searches. Am I believing }>too much of what I hear? }From: johnbr@atl.hp.com (John Bria) }Subject: Re: HELP - Hardware mirroring versus dbms mirroring }Date: 17 Aug 1994 21:54:52 GMT }X-Informix-List-Id: <news.8226> } }Just to add fuel to the fire: }HP's Logical Volume Manager will allow you to "stripe" non-array disks by }creating extents on a rotating basis across multiple drives. This may/may }not be advantageous as you develop a fragmentation strategy. } }If you use HP-UX mirroring, reads will be routed to the mirror copy if }the primary is busy. Under heavy disk loads, this is very advantageous.