Re: performance with many columns

Interprocess pipe size and task switching. An Oracle front-end communicates with an Oracle back-end through a pipe of one sort or another. The typical pipe size for tcp is 2K the typical size for on-machine pipes is 8K.

This leads to a very high rate of dialogue turnaround when you pass lots of long
rows from the front to the back.

Bottom line - communications in an
Oracle environment are more likely to
be bottled on turnaround time, not
on pure bandwidth.

You can increase the pipesize on local
connections by setting up tnsnames.ora
and listener.ora to have a maximum (usually) of 32K for the SDU (documented) and TDU (undocumented).

Do check, however, in v$session_wait and v$session_event for other possible lost time - there are usually other reasons for extreme amounts of time loss before turn-around time can be blamed. (Do a delta on v$session_event for the critical SIDs as your processes are running).

--

Jonathan Lewis
Yet another Oracle-related web site: http://www.jlcomp.demon.co.uk

Bart Blom wrote in message ...
>Some additional information:
>The data base block size is 8192; the record size is 973;
>data and indexes are on different tables spaces, on different
>SCSI controllers.
>All configuration parameters in Oracle have been set to the defaults
>for LARGE (in init.ora) or better.
>
>Most forms of access will always result in a while table scan.
>
>We have been running tests using view tables on the same actual table,
>with a greatly reduced number of columns (a process called folding) which
>will substantially increase performance.
>
>What I am looking for is information that can explain why large numbers
>of columns has such a detrimental impact on performance, and/or
>suggestions on how to alleviate this.
>
>thanks,
>
Received on Fri Jan 07 2000 - 10:52:22 CST