Thank you Jonathan.
Actually i posted two queries in this thread earlier (as below) one was
INSERT and Other was SELECT. Lothar pointed out in both of the cases the
execution time was spent in the query coordinator section and how the
complete execution path of the query was not happening in parallel. Some
parts were being serialized.
https://gist.github.com/oracle9999/ff2073c222398416c8095d109c233765
https://gist.github.com/oracle9999/618251c1e48b315dc70c73e157443773
In case of INSERT query as you see, the parallel execution was not
happening initially and it was enable_parallel_dml which was missing. But
after applying that hint, I got the below message in the Note section of
the query , which exposes the restriction which GTT is having in oracle.
Note
-----
* - PDML disabled because temporary tables with indexes used*
- parallel query server generated this plan using optimizer hints from
coordinator
The other thread I initiated to put this observation in a cleaner fashion,
and to see if any workaround is possible as we use global temporary tables
heavily in one of the customer applications and this restriction with
parallel execution for indexes does not feel good. And correct me if wrong,
I believe converting all the queries(noth INSERT/SELECT) into UNION ALL to
avoid this restriction is not a good idea and also it may not always be
possible I think.
On Sat, Jul 1, 2023 at 3:04 PM Jonathan Lewis <jlewisoracle@xxxxxxxxx>
wrote:
Append should be faster as compared to the other one (i.e. with justPeople here are expecting the query which runs using parallel(4) with
parallel(2)). But it's not happening that way. Want to understand why?
I see there's already a trail of notes about the index range scan not
running parallel, but I can't see any comment addressing the your question
about the APPEND hint. One possible explanation is that the index
maintenance mechanisms for "load as select" are different from those used
for "conventional load".
Unless things have changed in recent versions of Oracle:
Under conventional load each index is maintained in "real time", i.e.
as rows are inserted.
Under direct path load Oracle creates an index segment on the appended
table segment, then merges the original index with the new index segment
and writes a new index segment.
I'm sure I wrote a document about this once, but I can't find it at
present. The thing that's most likely to have changed is the way that
Oracle merges the newly created index segment with the existing index
segments - and there may be all sorts of variations anyway for partitioned
tables, parallel loads, and GTTs.
Depending on the number of indexes, the size of the table and its indexes,
and the volume of data loaded a conventional load might, therefore, do less
work than the direct path load.
Regards
Jonathan Lewis
On Sat, 24 Jun 2023 at 15:44, Pap <oracle.developer35@xxxxxxxxx> wrote:
Hello Listers,
This Oracle database version 19C and its exadata. We have a customer
query in which one of the executions is happening in parallel(2) and
another execution is happening in parallel(4) in two different places. The
query is exactly the same , just that the place in which it's getting used
with PARALLEL(4) it's also happening with the APPEND hint.
People here are expecting the query which runs using parallel(4) with
Append should be faster as compared to the other one (i.e. with just
parallel(2)). But it's not happening that way. Want to understand why?
And also in this query the step no which is coming as top contributor in
ASH is table access RFFT using index access path - RFFT_IX7. This is a big
table partitioned on column C_KEY. Other tables are small tables. So just
wondering if the higher parallel threads were not helping the query because
the index access is happening using BATCHED index range scan (i.e. TABLE
ACCESS BY LOCAL INDEX ROWID BATCHED). or are we misinterpreting the
execution plan here and something in the query is not making the index scan
to happen in parallel?
Below is the sql monitor details of the two plans. Below are two sql
monitors, in one case it has processed 2million vs other 3million , however
if we see the processing speed i.e. ~67 seconds for 3million vs 38 seconds
for 2million. The throughput seems to be more or less the same only. The
Parallel-8+ Append seems to be comparatively slower though, which is odd.
https://gist.github.com/oracle9999/618251c1e48b315dc70c73e157443773
Regards
Pap