Dontcheff

Archive for April, 2015|Monthly archive page

Table Temperature, Big Tables In-Memory and Automatic Big Table Caching

In DBA, Init.ora, Oracle database on April 3, 2015 at 16:10

If you have never heard of Oracle object temperature and how this relates to caching, in-memory and big tables, here is a short note perhaps worth reading.

And all this is purely Oracle database related, nothing to do with Oracle Lighting Products.

Oracle_light

Background: Oracle 12.1.0.2 came with 5 new init.ora parameters (actually 12 but the other 7 are In-Memory related):

1. common_user_prefix
2. dbfips_140
3. enable_goldengate_replication
4. exafusion_enabled
5. db_big_table_cache_percent_target

The first 4 are rather on-off type of parameters but the last one, db_big_table_cache_percent_target, is worth looking into.

A table is considered small, if the number of blocks in the segment is lower or equal than the value of the parameter _small_table_threshold. In 12c, this parameter defaults 2% of the buffers in the cache. To be more precise, it is 2% of _db_block_buffers. For details, check Jonathan Lewis’s Small tables. Another very good one is by Tanel Poder entitled Optimizer statistics-driven direct path read decision for full table scans.

Otherwise, we call the table big.

small_big_table

In 12.1.0.2, a DBA can enable full database caching with the following command: ALTER DATABASE FORCE FULL DATABASE CACHING;

Under normal running the Oracle database decides what data to cache in the buffer cache. If there is not enough room, data can be aged out of the cache. If Oracle determines that the buffer cache is big enough to hold the entire database it will cache all blocks.

However, seldom we have enough memory to cache the whole database. In this case, the parameter db_big_table_cache_percent_target can be used. And of course, for “big tables” only.

The default value is 0, so under normal circumstances automatic big table caching is not enabled. You can decide on your own if this is Poor Man’s In-Memory Caching but even if it isn’t the benefits are rather questionable. Let us see why.

We have 3 tables: SALES, OLD_SALES and VERY_OLD_SALES.

sales_objects

SALES and OLD_SALES are huge and SALES is in-memory, that is the in-memory option has been enabled for the SALES table. VERY_OLD_SALES is a rather small table but still considered big by Oracle as the small table threshold is 401 (2% of 20090):

select a.ksppinm name, b.ksppstvl value 
from x$ksppi a, x$ksppcv b 
where a.indx = b.indx 
and a.ksppinm like '%small_table%';

_small_table_threshold
401

select a.ksppinm name, b.ksppstvl value 
from x$ksppi a, x$ksppcv b 
where a.indx = b.indx 
and a.ksppinm like '%block_buffer%';

_db_block_buffers
20090

Let us see if the table is In-Memory enabled and what is the number of its blocks:

sales_tables

3 new dynamic views give us details on big table caching:
– V$BTS_STAT
– V$BT_SCAN_CACHE shows the parameters and status of the big table cache section.
– V$BT_SCAN_OBJ_TEMPS shows the active objects currently tracked by the big table cache.

As SALES is an IM-table (DATAOBJ# is 91985), regardless of how often you scan it, no buffers will go to the big table cache (I have set db_big_table_cache_percent_target to 50):

big_table_cache

If you do a full table scan on OLD_SALES, as it is way to big for the BT-cache (= Big Table Cache), no use either… I have tested and at least V$BT_SCAN_OBJ_TEMPS shows nothing.

However, when I do a subset scan on OLD_SALES, then the BT-cache is used as the buffers will fit.

Now, let is look at the temperature of the tables. Oracle assigned 2000 for the temperature of VERY_OLD_SALES (DATAOBJ# is 92326) while 1000 for the temperature of OLD_SALES (DATAOBJ# is 92322) which is only partially loaded into the BT-cache. After updating VERY_OLD_SALES, the temperature grows to 5000:

temperature

The more SQL & DML we do on the table, the higher the temperature gets. A detailed study by Mahmoud Hatem can be found here.

In short:

– It is not possible to control this feature on table basis, there are no attributes/parameters like KEEP or STORAGE clauses on table level.
– If we dedicate real memory just for such type of caching, then it will not always be used for mega huge tables that anyway will not fit, it will only be used for subset retrievals.
– The memory could be more useful for the IM option as benefits are much more clear and obvious.
– The BT-cache is populated also by the full index or range scans.
– CBO is not aware that segments are being cached.
– One cannot manually flush the BT-cache but it is gone after flushing the buffer cache.
– If the BT-cache is not big enough for several objects, then the ones with highest temperature are given the priority.

Big table cache: an optional, integrated portion of the database buffer cache that uses a temperature-based, object-level replacement algorithm instead of the traditional LRU-based, block-level replacement algorithm.

Advertisements