Jonathan, I hinted to use HASH join, and as you said it was lot faster -
returns in15 mins overall. Tried increasing the parallelism upto 32, but I
only gain 2 mins with higher degrees.
Here is the same document with the new statement towards the end:
https://drive.google.com/file/d/1hIUHbVfBxsvJz5zC3g5ZZGRg99MPz6p7/view?usp=sharing
Now I need to find out why it was not using the HASH in the first place
without hint and also see if I can make it even faster.
On Fri, Jul 10, 2020 at 4:57 PM Ram Raman <veeeraman@xxxxxxxxx> wrote:
Thanks Jonathan. There is no compression, but many columns dont have
values in them. I am bit puzzled by the naming convention, but that is a
different topic
What you said makes good sense. It doesnt make sense to access them by
nested loops. Here are the big table stats:
NUM_ROWS BLOCKS LAST_ANAL AVG_ROW_LEN CHAIN_CNT
------------------------ ------------ --------- ----------- ----------
1,738,075,641 40,228,648 24-JUN-20 371 0
Only one block size for the whole DB:
SQL> sho parameter db_block_size
NAME TYPE VALUE
------------------------------------ --------------------------------
------------------------------
db_block_size integer 8192
Partitions stats for the big table:
PARTITION SUBPARTIT PARTITION_COUNT STATUS
--------- --------- --------------- --------
HASH NONE 32 VALID
I did some calculation myself. If the query has to retrieve 14M rows from
the big table, given the average row length is 371 and hence rows per block
is about 22, it has to visit 634K blocks. Does not make sense to do nested
loop join for that. Let me try a HASH join hint
Thanks
Ram.
On Fri, Jul 10, 2020 at 2:07 PM Jonathan Lewis <jlewisoracle@xxxxxxxxx>
wrote:
Have I missed some posts on this thread ?
It looks to me as if your problem is simply a bad execution plan -
possibly due to bad stats.
The "inc" distinct count is about 1,700 but your total row count is 14M,
which means you're using an index to pick up about 8,000 randomly
distributed rows from a big table/ for each distinct value in the nested
loop driver. The optimizer thinks (for whatever reason) it will be about
1,600 rows fairly well clustered (about 8 rows per block accessed). At 14M
rows that's about 1.75M random I/Os - and there's little doubt that a lot
of them will collide and most of them will be real physical reads.
How many partitions in the table and what is the block size ? (It looks
like it might be 16KB blocks - and that might have an effect on
performance).
You're optimum plan might be a parallel hash join with a broadcast/none
distribution to minimise messaging between parallel execution processes.
One thing that might make the processing slow is your comment that you've
got 30 columns of varchar2(4000) - even though the average row size seems
to be in the region of 200 or 400 bytes - so maybe you've also got
compression in place. One feature of very long rows is that if the
declared length of the rows is greater than the size of a block Oracle will
probably fall back to single row processing on the insert. (I'd have to
check a combination of things to be sure, including the effect of the /*+
append */ hint). So it's almost certainly a good idea to pre-declare the
target table - ideally limiting the column sizes to get a row size less
than the block size.
Regards
Jonathan Lewis
On Fri, Jul 10, 2020 at 6:12 PM Ram Raman <veeeraman@xxxxxxxxx> wrote:
Thanks Andy. I am not sure about a covering index as the table has
around 50 columns, with 30 of them VARCHAR2(4000)! Not sure if creating a
covering index on such a table with 2B rows a good idea.
I did some tests and as you pointed out the count(*) comes back fast,
but when i do a SELECT of all columns instead of count from the big table
the query slows down to about 3 hours like the create table. I guess it is
time for new hardware.
On Wed, Jul 8, 2020 at 2:09 AM Andy Sayer <andysayer@xxxxxxxxx> wrote:
Hi Ram,
Covering index means that it covers every column that’s required by the
query (even the ones in your select). Judging by the cost in the plan, it’s
unlikely the index used contained more than one column. Covering indexes
remove the need for table access. You would need to weigh up the cost of
creating an index the size of the table and the reward (this query gets
faster).
Thanks,
Andrew
On Wed, 8 Jul 2020 at 03:25, Ram Raman <veeeraman@xxxxxxxxx> wrote:
Thank you Tim and Andy,
" besides making a covering index" Andy, there is a covering index
which makes the access of CLN_EVNT table by index, per the plan
On Tue, Jul 7, 2020 at 5:39 PM Andy Sayer <andysayer@xxxxxxxxx> wrote:
Tim,
"Load as select" means it's doing direct path load. The additional
step of "OPTIMIZER STATISTICS GATHERING" also only appears for direct
path
load.
We can also tell that the insertion into the table is done in
parallel as we are still using PCWP in the IN-OUT column, it only
serializes about the load.
Thanks,
Andy
On Tue, 7 Jul 2020 at 23:25, Tim Gorman <tim.evdbt@xxxxxxxxx> wrote:
Ram,
Perhaps it is serializing on the INSERT phase of the CTAS, keeping
in mind that the CREATE clause both creates the table and does the
insert?
Perhaps...
ALTER SESSION ENABLE PARALLEL DDL;
CREATE TABLE ... PARALLEL 8 SELECT /*+ full(ce)
parallel(ce,8) */ CE.* FROM ... ;
Unless you are required to perform CTAS in one statement, could you
perhaps split it into two steps, the first to create the table and the
second to populate it? So as an alternative...
CREATE TABLE .... SELECT * FROM IX.S_CLN_EVNT WHERE 1 = 0;
ALTER SESSION ENABLE PARALLEL DML;
INSERT /*+ APPEND PARALLEL(x,8) */ INTO REQ_27333_CLN_EVENT
SELECT /*+ full(ce) parallel(ce,8) */ CE.* FROM ... ;
This way you can use the APPEND hint to force direct-path INSERTs
which are far more efficient than conventional INSERTs; I'm not sure if
LOAD AS SELECT (HYBRID TSM/HWMB)" is direct-path or conventional? At
any
rate, with an APPEND hint you can be certain. You'll also get to
specify
the PARALLEL hint on the INSERT statement as well as the SELECT
statement,
to make sure of PX pipelining. Also, not sure if your session is
enabled
for parallel DDL or parallel DML, but if you break it into two
operations,
you will only need parallel DML.
Hope this helps?
-Tim
On 7/7/2020 2:48 PM, Ram Raman wrote:
Hi,
We are trying to create a table by pulling information from a big
table, the create statement is taking too long. However when I run
select
count(*) with the same query it comes back in seconds. While the CREATE
statement is running I see waits on parallel read almost all the time,
however, I do not see the same waits during select count(*).
Tried a few variations with parallelism degree in the Create
statement. Also tried inserting after truncating the table, used INSERT
with APPEND hint, same slow results. v12c. FORCE_LOGGING is not on at DB
level, inserted table in nologging mode. During the run, the IO on the
data
disks is about 50 to 70% util. 2 cores are consistently at 99-100%
(most of
them are %iowait at ~95%) with the other cores idle during this run the
last time I checked - I only checked the individual core usage this last
time it was run. I was checking average before. The wait events are the
same in all cases.
The attached link
<https://drive.google.com/file/d/1N8re2eAWLl0FHYdGs0c9I93e5ruoSNN6/view?usp=sharing>
has all the commands and the information. Any help is appreciated.
Thanks,
Ram.
--
--
--
--
