Re: Performance degradation question w/transactions

In article <+7jKSBANAxnzEwZS@smooth1.demon.co.uk>, David Williams <djw@smooth1.demon.co.uk> writes: > >Can anyone suggest what might be degrading the performance of my > >system so horribly in tests 1 and 3? Are there parameters that I > >can or should tune in my Informix configuration to avoid getting > >bogged down? > > Yes, stop commiting so often, this is not a realistic test as I > doubt you will get this many commits in a real world application. > however.... assuming that you do need to do this many tps, then take a look at: onstat -p # p = performance outout.... onstat -F # F = for what do I know? :-) Essentially you want to see what your %write cache is doing. I suspect that it's pretty low... but that's a small piece of the pie... Compare that information against what's in your online log file... look specifically for checkpoint duration. See if it's lasting a long time (> 3-5 seconds)... if it is, perhaps your physical log is *too* small. You may need to increase it because when it hits 75% full, a checkpoint will start causing your buffer cache to be flushed out to disk. (Informix' checkpoint is very traumatic, IMHO, because it *suspends* all writes while it's happening). You can verify if this is happening by comparing the checkpoint interval variable (CKPTINTVL) against the checkpoint activity that is logged in your online.log file. If the CKPTINTVL is greater than what is noted in the log, then you're more than likely getting a checkpoint because your PHYSFILE is too small and is filling... crank it a bit... You can reduce the checkpoint duration by tweaking the LRU_MIN_DIRTY and LRU_MAX_DIRTY parameters too. Set them lower and you'll have more LRU writes (monitored using onstat -F)... which in turn will reduce the number of chunk writes. -- Pablo Sanchez | wk: 415.933.3812| pg: 800.930.5635 -or- pablo_p@pager.sgi.com --------------+-----------------+-------------------------------------------- pablo@sgi.com ... when mailing me, place "not spam" in the Subject