You're testing with 1 to 8 CPUs and a parallel tablescan (or index fast
full scan) with a requested DOP of 1024, and finding that the actual degree
of parallelism appears to vary with the number of CPUs.
There are many factors that might be relevant - to start with
a) is this a PDB within a CDB, or non-container setup
b) have you collected system stats, or used the call to calibrate_io to
tell the optimizer something about the effectiveness of parallel query
c) have you allowed ALL the parallel-related parameter to default, or have
you set some of them explicitly
d) have you set the PROCESSES parameter explicitly, or do you allow it to
default
e) have you set the pga_aggregate_target to a non-zero value, what about
the sga_target or memory_target
f) how large is the table you're scanning in extents, and in blocks
g) how many data files are there in the database, and in the tablespace
holding the table
h) how many logical devices (that Oracle might be able to detect) is the
database / tablespace spread over
-- (f) might be relevant but I don't really think (g) and (h) are likely
to be relevant, though it's always worth checking.
When you run each experiment do you check ALL the parallel parameters each
time to see what has changed from previous tests.
Just to confirm my understanding - when you've said "sessions" above, you
mean sessions involved in the tablescan rather than the total number of
sessions execution the query which would be double that plus 1; hence the
"n_max" figures you're quoting that reach a maximum of 228 is the apparent
limit on your use of parallel server processes, not the degree of
parallelism of the query, which would have been half that.
Regards
Jonathan Lewis
On Fri, 27 Nov 2020 at 21:19, ahmed.fikri@xxxxxxxxxxx <
ahmed.fikri@xxxxxxxxxxx> wrote:
Hi Laurentiu,
I managed to write a function that returns the SID for the session that
read the block in the select statement. The function looks like this:
select count(distinct sid) from
table(my_function_accept_ref_curssor_and_get_exactely_the_sid_getting_the_row(CURSOR(SELECT
/*+parallel(t 1024) */ object_id from usr_dummy.big_tbl t)))
this returns:
number_of_sessions = max(floor(min(parallel_max_servers,n_max)/*2*),1)
And while executing the sql I also verified the number of sessions from
the v$session:
SELECT COUNT(DISTINCT SID) a FROM v$session WHERE status = 'ACTIVE' AND
sql_id = '3w8kg71wmsy4g';
and this returns (so two times the above value + 1)
number_of_sessions = 1+max(floor(min(parallel_max_servers,n_max)),1)
I got the formula using a script that I wrote. The script performs the
tests by changing the parallel_max_servers from 1 until 2 x n_max.
So I'm still trying to explain how does oracle choose these limits:
cpu (host) n_max
1 40
2 80
3 120
4 160
5 200
6 228
7 228
8 228
The background to my investigations is that after my client upgraded
Oracle from 11.2.0.4 to 12..1.0.2 a regression test failed (some
calculations are wrong, after setting the parallel_max_servers to 1, the
calculation is again correct).
It looks like a bug in oracle.
Best regards
Ahmed
-----Original-Nachricht-----
Betreff: Re: parallel_max_servers and the number of sessions involved in a
SQL
Datum: 2020-11-27T21:35:21+0100
Von: "Laurentiu Oprea" <laurentiu.oprea06@xxxxxxxxx>
An: "ahmed.fikri@xxxxxxxxxxx" <ahmed.fikri@xxxxxxxxxxx>
Hello Ahmed,
If your target is to limit the dop you have to either:
-> mention DOP number in parallel hint : /*+ parallel(n) */
-> explore PARALLEL_DEGREE_LIMIT
Also DOP of 8 don`t translate into 8 sessions. You have involved the
coordinator, producers/consumers, parallel servers already allocated (you
might experience DOP downgrade). (you can check the sql monitor report for
comprehensive details )
You might want to avoid relying on autodop algorithm.
În vin., 27 nov. 2020 la 21:48, ahmed.fikri@xxxxxxxxxxx <
ahmed.fikri@xxxxxxxxxxx> a scris:
Hi all,
I'm trying to understand how Oracle defines the limit on the number of
sessions involved in a SQL statement (a select on big table using a
parallel hint)
I ran some tests and I found the following:
number_of_sessions =
greatest(floor(min(parallel_max_servers,n_max)/2),1)
where
cpu (host) n_max
1 40
2 80
3 120
4 160
5 200
6 228
7 228
8 228
for my test I used 12.2.0.1 installed on a virtual box (the host machine
has 12 cpus)
can someone explain the n_max/cpu relationship?
Best regards
Ahmed
