Re: Page cleaners and raids

Andrew Hamm wrote: > > Mark D. Stock wrote in message <91bm23$d5q$1@news.xmission.com>... > > > >Andrew Hamm wrote: > >> > >> On that subject, (and sorry to diverge from what you may understand, > Debbie) > >> can someone clear up a question I've been wanting to ask? With the > >> split-reads offered by Informix mirroring, does it split the addresses > thru > >> the middle of the total chunk size, or thru the middle of the > high-watermark > >> of the currently allocated space in the chunk? Seems to be it should take > >> the watermark into consideration, cos if you allocate a 1G space but > >> currently only load 200M, then you ain't gonna get split reads. > > > >My understanding is that it has nothing to do with chunk size, but > >rather the number of pages to be written. They are sorted by chunk, then > >into disk order, so that the first half can be written to the primary > >chunk and the second half to the mirror chunk. > > > >This algorithm is very simple, but may well have changed now in version > >9.x. It is unlikely that Informix mirroring will beat H/W mirroring that > >can do load balancing because other access to those disks is taken into > >consideration. > > > Emm, I was asking about split reads, page 23-9 of the admin guide (V7). > > You sound like u are saying that writes are also split, but I didn't think > that was possible because both chunks MUST be written. Are you implying > there is some sort of priorisation going on along the lines of split reads? > Can you give a page number? Now that would imply that I read the manual. ;-) > I'm inclined to say the opposite about mirroring winners - I'm talking about > a system that is principally doing Informix only (and running the apps, > which don't do any serious I/O through file systems). Since the engine > submits the two writes asynchronously, then the load balancing in the write > queues will allow the two disks to be doing different things. What we need > is hard numbers. Can you justify with numbers? From what I've read about > RAID, some types will be definitely slower than informix mirroring, but I > can see that HW mirroring could have the most chance of out-performing IFX > mirrors, depending on quality of the system. What do you think about O/S > software mirroring? Crap? S/W mirroring is the only option in certain situations, like when you don't have a RAID controller. It adds an extra layer and thus degrades performance. It certainly depends on how you are using the disks as to which option gives best performance, but it is rare for Informix mirroring to outperform H/W mirroring. No I don't have any figures, and even if I did, you shouldn't believe them. They may be meaningless for your system and lead you to implement the wrong solution. Cheers, -- Mark. +----------------------------------------------------------+-----------+ | Mark D. Stock mailto:mdstock@mydas.freeserve.co.uk |//////// /| | http://www.informix.com http://www.informixhandbook.com |///// / //| | http://www.iiug.org +-----------------------------------+//// / ///| | |This email will self-destruct in |/// / ////| | |10 sec. If you received this email |// / /////| | |in error, sorry about the mess. |/ ////////| +----------------------+-----------------------------------+-----------+

Mark D. Stock wrote in message <91gt1q$46r$1@news.xmission.com>... > >S/W mirroring is the only option in certain situations, like when you >don't have a RAID controller. > I'd call Informix mirroring a SW mirroring, but at least it has the smarts to understand it's own needs, rather than some assumed average, lowest common denominator that O/S SW mirroring must necessarily assume. From that theory, I'd totally ignore O/S mirroring and stick with either IFX mirroring or hardware mirroring if it can be proven to be faster in production (for certain quality hardware?) I think the biggest trouble we all have with getting definitive answers is the cost in time and disruption of performing hard tests to prove one way or the other. We all end up with pet theories about so many things to do with the tuning of the engines. I'm thinking of the story of the Elephant and the Blind Men right now... >No I don't have any figures, and even if I did, you shouldn't believe >them. They may be meaningless for your system and lead you to >implement the wrong solution. > I believe that significant results can be "ported" from application to application. I mean, if someone can say "hey, the hardware mirroring on HP using controller X with 300 concurrent database connections" outperforms IFX mirroring by a factor of 2:1 for reads and 3:1 for writes, then I'd gamble that those figures are definitely significant to all users of HP+X. Any contradictions to this? Note that I am attempting to quantify the machine load by naming the number of users; but different applications no doubt have wildly varying efficiencies. Perhaps we could perhaps talk about log throughput as a coarse measure of activity, or gross read and write statistics during an average hour. Whatever the quantification of the application, I'm willing to rely on numbers as a guideline.