Re: very (I repeat) slow DBWR

Gerard H. Pille wrote in message <01be96cf$41038920$390120c0_at_GHP.creyfs.be>...
>The data-files of the tablespace being loaded via import, are on one
>SSA-disk. This disk also contains the log of the Journaled File System.
>The DBWR(s) show(s) activity of around 70-80.
>hdisk3_b and hdisk4_b each contain a copy of the JFS-log, so less than 80
>Kb were written to the data-files in 5 seconds.[...]
>I started with 4 DBWR-processes with use_async_io = false, then switched to
>async_io which resulted in the DBWR doing all the work (sic) and the 4
>child-processes idling. Same lack of performance.

Slaves consume small amount of CPU because they perform only I/O. All cpu-intensive work (like buffer lists scans) is done by master (dbwr). I suspect some problems related to latches preventing dbwr from effective maintaining dirty buffer and free buffer list. BTW, is it multiprocessor machine ?
Sure, some problems may be related to particular RDBMS release. Let's know about your RDBMS version (not dying processes, but you should try using shutdown immediate instead of shutdown normal; ^C; shutdown abort) and buffer cache size.
Some problems may be easily found via utlbstat/utlestat report.

--
Piotr Kolodziej pkol_at_otago.gda.pl
Just my private opinion. Received on Wed May 05 1999 - 04:03:55 CDT