_very_large_object_threshold
Posted by Игорь Усольцев on June 4, 2013
From Oracle 11.2 choice between serial direct path read and buffered read (through db buffer cache SGA) for INDEX FAST FULL SCAN (IFFS) operation depends on parameter:
SQL> @param_ _very_large_object_threshold NAME VALUE IS_DEF DSC ---------------------------- ----- -------- ----------------------------------------------------- _very_large_object_threshold 500 TRUE upper threshold level of object size for direct reads
The correct parameter meaning is: the percent of buffer cache size (more precisely % of _db_block_buffers value). If index size exceeds this threshold than this index is considered as “huge”, and for IFFS operation will be used direct path read. Otherwise IFFS will use standard buffered reads – exactly as Sayan Malakshinov described in Example of controlling “direct path reads” decision through SQL profile hints (index_stats/table_stats). In the same note Sayan described hint INDEX_STATS(“OWNER”.”TABLE_NAME”, “INDEX_NAME”, scale, blocks=X), which can be often noticed for SQL Profiles usage
Simple tests show _very_large_object_threshold influence on direct path read usage for full table scan, together with event 10949
Sayan test schema:
drop table xt_iffs purge
/
create table xt_iffs as
with gen as(
select level a,mod(level,10) b,lpad(1,50,1) c
from dual
connect by level<=1e3 )
select gen.* from gen,gen gen2
/
exec dbms_stats.gather_table_stats('','XT_IFFS')
create index ix_iffs on xt_iffs(a,b)
/
On Windows x86_64 platform:
SQL> @inst
INST_ID INSTANCE_NAME VERSION PLATFORM_NAME
------- -------------------- ---------- ----------------------------
1* orcl1123 11.2.0.3.0 Microsoft Windows x86 64-bit
SQL> @param
NAME VALUE
------------------------------------------ ---------
_very_large_object_threshold 500 -- default value
_db_block_buffers 30442 -- on instance startup in blocks
db_cache_size 0 -- ASMM
__db_cache_size 260046848 -- current buffer cache size in bytes
SQL> select blocks, trunc(bytes/1024/1024) as MB, trunc(blocks/30442*100,2) as "threshold%" from user_segments where segment_name = 'IX_IFFS';
BLOCKS MB threshold%
---------- ---------- ----------
2560 20 8.4 -- our test index size is 20 MB, which is 8.4% of _db_block_buffers
SQL> SELECT SUM(A) FROM XT_IFFS T;
SUM(A)
----------
500500000
1 row selected.
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
logical read bytes from cache 20447232 -- query with IFFS uses buffer cache SGA
...
physical reads cache 2485
physical reads 2485
free buffer requested 2485
...
index fast full scans (full) 1
...
Session Wait Events
NAME WAITS TIME_MS TIMEOUTS AVG_WAIT_MS
---------------------------------------------------------------- ------------ ------------ ------------ -----------
SQL*Net message from client 2 66604 0 33301,8
db file scattered read 53 132 0 2,5 -- top non-idle wait event confirm buffer cache IO usage
SQL*Net message to client 2 0 0 0
-- Now we artificially make index HUGE
SQL> alter session set "_very_large_object_threshold"=8;
Session altered.
SQL> SELECT SUM(A) FROM XT_IFFS T;
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
...
physical reads 2486
physical reads direct 2486
...
index fast full scans (direct read) 1 -- query uses serial direct read
index fast full scans (full) 1
-- if we a bit increase parameter - index again becomes SMALL
SQL> alter session set "_very_large_object_threshold"=9;
Session altered.
SQL> SELECT SUM(A) FROM XT_IFFS T;
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
logical read bytes from cache 20447232 -- and again IFFS uses buffer cache
consistent gets from cache (fastpath) 2496
consistent gets 2496
session logical reads 2496
consistent gets from cache 2496
...
index fast full scans (full) 1
In last tests there is not clear how to calculate _very_large_object_threshold value correctly, because this percent calculation based on __db_cache_size insignificantly differs from the same of _db_block_buffers value:
SQL> select blocks, trunc(bytes/1024/1024) as MB, trunc(blocks/(260046848/8192)*100,2) as "threshold%" from user_segments where segment_name = 'IX_IFFS';
BLOCKS MB threshold%
---------- ---------- ----------
2560 20 8.06
But dynamic (on fly) increase of current db_cache_size value (by ASMM usage!) can exactly shows that discussed threshold% is calculated by Oracle using _db_block_buffers value:
SQL> alter system set db_cache_size=300000000;
System altered
SQL> @param_ db_cache_size
NAME VALUE
------------------------------------------ ---------------------
db_cache_size 301989888
__db_cache_size 301989888
SQL> select * from v$sgainfo where name = 'Buffer Cache Size';
NAME BYTES RESIZEABLE
------------------------------------------ ---------- ----------
Buffer Cache Size 301989888 Yes
SQL> select blocks, trunc(bytes/1024/1024) as MB, trunc(blocks/(301989888/8192)*100,2) as "threshold%" from user_segments where segment_name = 'IX_IFFS';
BLOCKS MB threshold%
---------- ---------- ----------
2560 20 6.94 -- now index size is 6.94% of db_cache_size
SQL> @param_ _db_block_buffers
NAME VALUE
------------------------------------------ ---------------------
_db_block_buffers 30442 -- not changed from startup, as expected
-- and now index can be considered as HUGE if Oracle calculation is based on _db_block_buffers value,
-- (more than 8%), while index is SMALL relatively db_cache_size - less than 7%:
SQL> alter session set "_very_large_object_threshold"=8;
Session altered.
SQL> alter system flush buffer_cache;
System altered
SQL> alter system flush shared_pool;
System altered
SQL> SELECT SUM(A) FROM XT_IFFS T;
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
...
physical reads 2486
consistent gets direct 2486 -- now used direct read, which proves that threshold% is calculated from _db_block_buffers
physical reads direct 2486
...
index fast full scans (full) 1
index fast full scans (direct read) 1
Setting event 10949 don’t affect on serial direct path read choice for INDEX FAST FULL SCAN – which is quite expected from short event description 🙂
$ oerr ora 10949 10949, 00000, "Disable autotune direct path read for full table scan" // *Cause: // *Action: Disable autotune direct path read for serial full table scan.
And practical tests:
SQL> @param_ _very_large_object_threshold NAME VALUE ------------------------------------------ ----- _very_large_object_threshold 500 SQL> alter session set events '10949 trace name context forever, level 1'; Session altered. SQL> SELECT SUM(A) FROM XT_IFFS T; Session Statistics NAME DELTA ---------------------------------------------------------------- ------------ logical read bytes from cache 20447232 ... consistent gets from cache 2496 consistent gets from cache (fastpath) 2496 ... index fast full scans (full) 1 SQL> alter session set "_very_large_object_threshold"=8; Session altered. SQL> SELECT SUM(A) FROM XT_IFFS T; Session Statistics NAME DELTA ---------------------------------------------------------------- ------------ ... index fast full scans (direct read) 1 index fast full scans (full) 1 ... Session Wait Events NAME WAITS TIME_MS TIMEOUTS AVG_WAIT_MS ---------------------------------------------------------------- ------------ ------------ ------------ ----------- SQL*Net message from client 3 27492 0 9164,1 direct path read 30 30 0 1 ...
More interesting is fact that for full table scan operation event 10949 can be used for direct path read usage choice together with _very_large_object_threshold (as table segment size percent of _db_block_buffers value) for full table scan. But this dependence not so simple as for INDEX FAST FULL SCAN:
SQL> select blocks, trunc(bytes/1024/1024) as MB, trunc(blocks/30442*100,2) as "threshold%" from user_segments where segment_name = 'XT_IFFS';
BLOCKS MB threshold%
---------- ---------- ----------
9216 72 30.27 -- segment size is about 30% of buffer cache
SQL> alter session set events '10949 trace name context forever, level 1';
Session altered.
SQL> SELECT /*+ FULL(T)*/ SUM(A) FROM XT_IFFS T;
------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 4 | 2416 (1)| 00:00:29 |
| 1 | SORT AGGREGATE | | 1 | 4 | | |
| 2 | TABLE ACCESS FULL| XT_IFFS | 1000K| 3906K| 2416 (1)| 00:00:29 |
------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8749 consistent gets
0 physical reads -- direct read was not used
...
-- as we did for index, trying to make our table HUGE:
SQL> alter session set "_very_large_object_threshold"=30;
Session altered.
SQL> SELECT /*+ FULL(T)*/ SUM(A) FROM XT_IFFS T;
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
logical read bytes from cache 72310784
...
consistent gets from cache (fastpath) 8809
consistent gets from cache 8809 -- still reads over cache
...
table scans (long tables) 1
-- and a little more:
SQL> alter session set "_very_large_object_threshold"=26;
Session altered.
SQL> SELECT /*+ FULL(T)*/ SUM(A) FROM XT_IFFS T;
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
...
consistent gets 8743
consistent gets direct 8740
...
physical reads 8740
...
physical reads direct 8740
...
table scans (long tables) 1
...
table scans (direct read) 1 -- direct path read
...
Session Wait Events
NAME WAITS TIME_MS TIMEOUTS AVG_WAIT_MS
---------------------------------------------------------------- ------------ ------------ ------------ -----------
SQL*Net message from client 2 15199 0 7599,3
direct path read 144 331 0 2,3 -- and appropriate wait event
...
– furthermore there is not (often noticed in different sources) accordance between _very_large_object_threshold and table size in Megabytes (72 MB), imho. Moreover this would be very strange to use parameter with unit which depends on SQL Plan operation
IFFS tests on Linux x86_64 platform with significantly increased segment size confirm previous parameter definition, and obviously demonstrate an absence of any accordance between _very_large_object_threshold and segment size (in Megabytes). And besides these tests show how direct read usage speed up simple query execution almost twice!:
SQL> @inst
INST_ID INSTANCE_NAME HOST_NAME VERSION PLATFORM_NAME
------- -------------- -------------- ---------- ----------------
1* INST1 db1.domain.net 11.2.0.3.0 Linux x86 64-bit
SQL> @param
NAME VALUE
------------------------------------------ -----------
_db_block_buffers 3780500
db_cache_size 0 -- ASMM
__db_cache_size 34762391552
db_block_size 8192
SQL> create table xt_iffs as
2 with gen as(
3 select level a, lpad(1,50,1) c
4 from dual
5 connect by level<=1e4 -- table with increased segment size
6 )
7 select gen.*
8 from gen,gen gen2
9 /
Table created.
SQL> exec dbms_stats.gather_table_stats('','XT_IFFS')
PL/SQL procedure successfully completed.
SQL> create index ix_iffs on xt_iffs(a);
Index created.
SQL> select blocks, bytes/1024/1024 as MB from user_segments where segment_name = 'IX_IFFS';
BLOCKS MB
---------- ----------
213022 1664.23 -- index size significantly exceeds _very_large_object_threshold value
SQL> select round((213022/3780500)*100,2) as "threshold%" from dual;
threshold%
----------
5.63
SQL> alter session set "_very_large_object_threshold"=6;
Session altered.
SQL> SELECT SUM(A) FROM XT_IFFS T;
Elapsed: 00:00:38.06 -- IFFS over buffer cache twice slower
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
logical read bytes from cache 1715216384
physical read total bytes 1715159040
physical read bytes 1715159040
...
free buffer inspected 212179 -- for cache blocks reading
...
physical read IO requests 1645
physical read total multi block requests 1639
...
index fast full scans (full) 1
...
Session Wait Events
NAME WAITS TIME_MS TIMEOUTS AVG_WAIT_MS
---------------------------------------------------------------- ------------ ------------ ------------ -----------
SQL*Net message from client 2 73827 0 36913,7
db file scattered read 1643 23674 0 14,4
db file sequential read 2 14 0 7,1
SQL*Net message to client 2 0 0 0
-- with decreased threshold down to 5%:
SQL> alter session set "_very_large_object_threshold"=5;
Session altered.
SQL> SELECT SUM(A) FROM XT_IFFS T;
SUM(A)
----------
5.0005E+11
Elapsed: 00:00:16.47 -- direct read more than twice quicker!
Session Statistics
NAME DELTA
---------------------------------------------------------------- ------------
physical read total bytes 1715167232
...
consistent gets direct 209371
physical reads direct 209371 -- direct read
...
index fast full scans (direct read) 1
...
index fast full scans (full) 1
Session Wait Events
NAME WAITS TIME_MS TIMEOUTS AVG_WAIT_MS
---------------------------------------------------------------- ------------ ------------ ------------ -----------
direct path read 795 8218 0 10,3 -- with one appropriate wait event
SQL*Net message from client 2 6492 0 3245,9
events in waitclass Other 2 3 0 1,5
enq: KO - fast object checkpoint 2 0 0 0,2 -- and another appropriate wait event
Oracle mechanics 
Oracle SQL | Example of controlling “direct path reads” decision through SQL profile hints (index_stats/table_stats) said
[…] UPD: btw previously I wrote that _very_large_object_threshold is a percentage of _db_block_buffers and Igor Usoltsev have tested and confirmed my guess […]
Oracle SGA 값을 증가 시킬 때 발생 장애 원인 – DBA의 정석 said
[…] ? […]