Re: Checkpoint duration increase with 7.22.UC1

In article <Pine.BSF.3.91.970520102915.23101B-100000@kilimanjaro.dallas. herald.net>, Richard Stanford <richards@herald.net.REMOVE> writes >In article <5lak28$ssn@cssun.mathcs.emory.edu>, Kernoal Stephens ><kernoals@earthlink.net> arranged the electrons to spell out: > >> With 7.13.UC2 my checkpoint duration was averaging 4.5 seconds. >> I used the same onconfig settings with 7.22.UC1 and my average >> checkpoint duration >> is 8.4 seconds. > >> What has changed to cause the checkpoint duration to increase? > >> ONCONFIG FILE >> ##################### >> > >> BUFFERS 172032> >This may have something to do with it. When we went to 7, we noticed a >similar problem. Under 5, having buffers set as high as we had memory >never gave us any grief, but there's a much more measurable correlation >between cache size and checkpoint duration under 7. We cranked buffers >down until a performance loss was found (to a much lower setting than we >expected) and then up a little. This got us significantly shorter >checkpoints. >I don't know if this is your situation or not, but it may be worth a try. > This is because under certain conditions 7.x can use more memory than 5.x. This means the BUFFERS setting may need to be REDUCED to decrease memory requirements and therefore stop paging/swapping happening. >As an aside, we got an EMC disk subsystem in to play with, and were able >to drag buffers down to, oh, 1000 or so. The danged thing got faster >every time we lowered buffers (and allowed the subsystem to cache, with >all of the attendant network overheads). This leads me to believe that >Informix doesn't cache as well as it could, but that was (in all fairness) >only one test. > Informix caches very well, onstat -p should should read caching (rache) at 98=99% and writing caching (wache) > 80&. If not then you have too few buffers. >DOH! > >I just reread your message and noticed that you were talking about >7.13 -> 7.22. My apologies. Still, this may be useful to other people >doing 5 -> 7, or may help you lower checkpoint duration (although it isn't >the original problem) so I'll send it anyway. > True... Any chance of an onstat -a from the person who is having problems? Yes I know it will be large (I'm just about to see one to Informix as I've hit a bug in OWS 7.20 - when you exceed the LOCKS parameter OWS crashes with an internal error mt.c, line....). >-Richard -- David Williams