Have a look at the DELTA_READ_IO_BYTES column (and related columns) in ASH
to determine what is using your IO bandwidth (there might be more relevant
columns for exadata). Those direct path read temp waits could easily be
hammering you here but we should get some numbers first. It’s worth
graphing out total IO per second and average response times for your
important data intensive queries to see at what level it starts to drop off
a cliff.
I don’t think the difference is related to foreign key validation - that
would give you extra IO which you would see on the monitor reports.
Hope that helps,
Andrew
On Fri, 2 Apr 2021 at 13:23, Eriovaldo Andrietta <ecandrietta@xxxxxxxxx>
wrote:
Hi Pap,
If you have a FK on the deleted table, check if the column of the FK in
the source table is indexed.
Regards
Eriovaldo
Em qui., 1 de abr. de 2021 às 07:58, Pap <oracle.developer35@xxxxxxxxx>
escreveu:
Hello all, We are using version 12.1.0.2.0 of Oracle exadata X3 machine.
Seeing variable performance for a DELETE query even though the plan is the
same. Though the number of rows are closely the same, we are seeing the
runtime is increasing from ~15minutes to ~30minutes. As its version 12C, so
I am able to fetch the sql monitor from these past two executions as
attached here. But i am not seeing any information in the activity detail
section of the report.
However I fetched the activity data from dba_hist_active_sess_history and
it says the "cell multiblock physical read" is quite high for the slow
execution. And also the objectid is pointing to the table partition which
means it was also not on UNDO. So does it mean that we must be suffering
from overall system slowness which impacted this query because of slow
overall response time for "cell multiblock physical read" in database level?
DELETE /*+ full(TAB1) */ FROM TAB1 TAB1 WHERE ETC = 'XX' AND ID = 3 AND
OTCD = 'YY' AND PART_DT = :B1
Select event, current_obj#, count(*) from dba_hist_active_sess_history
group by event, current_obj# order by count(*) desc;
*For faster run :-*
EVENT CURRENT_OBJ# COUNT(*)
cell multiblock physical read TAB1 73
TAB1 5
gcs drm freeze in enter server mode TAB1 1
*For slower run:-*
EVENT CURRENT_OBJ# COUNT(*)
cell multiblock physical read TAB1 170
TAB1 14
