Re: Mirroring

"Art S. Kagel" <kagel@bloomberg.com> wrote: >Jay Aymond wrote: >> >> I am adding mirroring to my Informix Online 5.x installation. Right now, I have >> Cleaners set to 2 for my two primary disk drives. Once I have enabled >> mirroring, should I change the number of Cleaners from 2 to 4 since I will have >> more disk drives, even though two of them are mirrored drives? > >If you want to use the formula of one cleaner per drive, yes. I assume >that tbstat -F shows most of your activity as 'Chunk Writes'? If so, >and this is primarily a query database, then this scheme may work best. > >BUT, if your instance is a high update transaction server, it tends to >get better database performance by reducing MAX_DIRTY and MIN_DIRTY and >increase the number of cleaners to match the number of LRU queues, >which should be 16, 32, or 64 depending on how large you buffer pool is. >This will result in 2/3 or more of your buffer flush activity being LRU >Writes and Idle Writes. The trade off is that the Informix engine will >have a greater impact on the system's I/O subsystem and not be as good >at 'Playing Nicely' with non-database applications. > >I know what the manuals say (one cleaner per disk and fewer LRUs is >better). The manuals are WRONG. The Informix benchmark people long >ago proved that and I have verified it myself. If your goal is to best >share the machine with other apps and servers and minimize impact on the >OS and hardware, follow the manuals. If you want maximum transaction >throughput, for BUFFERS=100000 go with 32 LRUs, 32 cleaners, >MAX_DIRTY < 10, MIN_DIRTY < 5. Adjust as appropriate. In this scheme >ignore the fact that you have mirrors, there will be plenty of free >cleaners to flush to the mirrors during Idle Write wake ups and LRU wake >ups and if your activity level is so high that there are still >significant numbers of dirty pages at checkpoint time you know that >there >will be more cleaners than disk so... > >With this scheme checkpoints are MUCH shorter and applications rarely >have >to wait for page flushes during checkpoints. I run my busiest servers >with >LRU_MAX_DIRTY=3 and LRU_MIN_DIRTY=0 90% LRU Writes, these puppies fly! > >Art S. Kagel Thanks for the info ---- I will give it a try. Is there some rule for defining how many LRU queues to have for a certain number of buffers? *************************************************************** Jay Aymond United Life & Annuity Insurance Company Email: jay@ulife.com ***************************************************************