Re: HP9000+LVM+Online

> > Andy > > I got the impression that Mike's system only had one chunk (therefore > one page cleaner) which was striped underneath using LVM. Disaster Boulevard for starters. The whole idea is to get as many page cleaners as possible (without putting them into contention with one another), writing to as many disks as possible in parallel. After all, parallelisation is the message behind the whole of Informix's current performance-boosting sales-hype message. And it happens to be true. Put one page cleaner on a five-disk system and you've already got your checkpoint performance down to 20% of what it might have been. > Note that LVM > is pure striping where one logical write equates to one physical write > therefore it doesn't suffer the write penalty found on RAID. > > In general, this would (theoretically) quadruple read throughput i.e. > given a long random list of read requests, it would get through them > four times quicker. Sorry, I just don't see how you arrive at that. AIUI a well-distributed chunk design can do anything you can do with LVM. > > For writes, the application inserts 250,000 times (20%) and updates 1M+ > times (80%). Again, LVM would mean that updates are quicker because > reading the record + index leafs would be faster. Nope. Don't see why. At least not over and above what you could do with standard OnLine facilities. > For sorted writes, you will still get a sorted write on each of the > underlying disks. e.g. if we have a 2 disk LVM, and Online writes > pages 1 2 3 4, this would translate to disk1page1, disk2page1, > disk1page2, disk2page2. As you can see, the actual physical write to > each disk is sorted. But you are halving the number of page cleaners you could configure. In this example, you'd merely be cutting write performance by 50%. I can't see the point in doing that. While you're writing to disk 1, disk 2 is twiddling its thumbs, and vice versa. How could this possibly be better than having two page cleaners happily hacking away at the two devices in parallel? > The only thing I'm concerned about is that some LVM products don't do > synchronous writes i.e. it buffers them. If this is the case on HP then > beware, because a checkpoint would not actually write to disk. Not only that, but neither would unbuffered log writes. Which after all is the crux of Informix's "guaranteed write" strategy. So bulletproof fault tolerance is instantly down the toobs. Aren't you glad Slough isn't a bit nearer Bristol, David ;-) . (Sorry folks, private exchange there) I rest my original case as I've *still* seen nothing here to disprove it:- LVM and RAID: Don't do them, kids. akent@cix.compulink.co.uk (Andy Kent) -------------------------------------