Re: Large SMP/UP Performance Deltas

In article <5032dr$pcp@rainbow.rmii.com>, Craig Hollenbaugh <csh@rainbow.rmii.com> writes >We're currently in the processes of benchmarking SMP performance, >against UP performance for our applications and have hit a wall >in terms of performance. (Boxes being tested are an 8-way R/S 6000 >and a 59H - fast UP). > >Problem is our application runs an order of magnitude slower (25 minutes >vs 2 minutes) on the SMP box. What I need help understanding is an >alarmingly high number of bufreads on the SMP box (10 million vs >1 million on the UP) running the same application on a database that was >"onunloaded" from the UP to the SMP. > >Interesting tidbits: UP - AIX 3.2.5 / Informix 7.10.UC1 > SMP - AIX 4.1.4 / Informix 7.12.UC1 >Disks are identical, SMP has 3x real memory, and both databases >are fragmented by expression. >Indexes verified with oncheck >Application is single threaded and one copy is running. > >My current tongue-in-cheek explanation is 8processors = 8times >the number of reads :-}, however I'm looking for another theory. > >TIA > >Craig Hollenbaugh csh@kru.com >DBA never wanna be Look in $INFORMIXDIR/etc/$ONCONFIG on the SMP Box - bet OPTCOMPIND is set to 2. Set it to 0 as otherwise you can probably getting dynamic- hash joins rather than indexed reads. i.e. the SMP Box is not using indicies. 7.12 has NEW default of 2 for OPTCOMPIND. Informix recommend changing it to 0 ASAP. With OPTCOMPIND 2 we were getting session crashes with supurious errors and (I think) incorrect results. PS I'm sure the default under 7.12 is 2 as I've just got in from work after setting up a test 7.12.UC1 instance at work!! -- David Williams