Hi Mates,

Today i will discuss about the Database In-Memory option of Oracle 12c databases. I know i am little late to add about this feature but recently I’ve implemented the option for one of my customer in Latin Americas region and there I’ve got more exposure to understand it better and implement the feature.

So, lets start with most probably my last post of the year!

Recently i have got a chance to test and to benchmark the performance gains in one of out Pre-Prod environment. This feature applies on Tablespaces, Tables, MViews, (Sub) Partitions (Except objects owned by SYS, SYSTEM and SYSAUX). To understand the feature i would first like to shed some light on two of the ways how Oracle stores tables on both Disk and Memory using conventional ‘Row Format’ and with the all new In-Memory ‘Column’ format.

Row Arrangement: Is the same old traditional method to store data in row formats. This is best for OLTP systems as queries runs faster with this approach as it quickly fetches all of the columns in a record.
Column Based Arrangement: This way it stores records in a separate column store. This proves good for OLAP systems where large set of data is chosen but only for few number of columns.

Hence, based on above two methods or data arrangements in database it is clear that the row based method is best for DMLs and column based arrangement is good when selecting large portion of data, so both of the two methods have their own respective pros and cons. But starting from Oracle 12.1.0.2 we have the all new feature of ‘DB In-Memory’ which use best of both the approaches. I mean it uses both row and column arrangements to keep data in memory. Our smart optimizer automatically knows which query to route as per the workload (OLTP and for Analytical processing).

The In-Memory feature uses the IM Column store which is a new occupant of the SGA (In-Memory Area). This In-Memory Area is adjusted by a parameter INMEMORY_SIZE.

As far as IM Column Store it is filled by the information collected by worker processes e.g. w001, w002 etc.; each of the worker process updates the IM Compression Units or IMCUs.

So, enough the background and internals involved, lets jump to the real work and check how this thing practically works.
First check the minimum comparability of the database

COMPATIBLE = 12.1.0.0.0

Lets first enable the in-memory column store. There are many of the related parameters and are given below.

NAME                                 TYPE        VALUE
------------------------------------ ----------- ------------------------------
inmemory_adg_enabled                 boolean     TRUE
inmemory_clause_default              string
inmemory_expressions_usage           string      ENABLE
inmemory_force                       string      DEFAULT
inmemory_max_populate_servers        integer     0
inmemory_query                       string      ENABLE
inmemory_size                        big integer 0
inmemory_trickle_repopulate_servers_ integer     1
percent
inmemory_virtual_columns             string      MANUAL
optimizer_inmemory_aware             boolean     TRUE

Before i show next logs on how to enable it, i would first like to show you the database startup details with no In-Memory enabled.

SQL> startup
ORACLE instance started.

Total System Global Area 1191182336 bytes
Fixed Size                  8620032 bytes
Variable Size             771753984 bytes
Database Buffers          402653184 bytes
Redo Buffers                8155136 bytes
Database mounted.
Database opened.
SQL>

Okay now lets enable it!
Now a question might be asked about the criterion of adding objects to the pool. So in my opinion the best candidates would
be — Very hot data, large segments (not less than 1MB), collect these stats from sources like Oracle segment statistics, number of column scans, AWR reports etc. Analytical queries etc.

SQL> alter system set inmemory_size=250m scope=spfile;

System altered.

SQL> shut immediate
Database closed.
Database dismounted.
ORACLE instance shut down.
SQL> startup
ORACLE instance started.

Total System Global Area 1191182336 bytes
Fixed Size                  8620032 bytes
Variable Size             822085632 bytes
Database Buffers           83886080 bytes
Redo Buffers                8155136 bytes
In-Memory Area            268435456 bytes
Database mounted.
Database opened.

Now i will try to load some stuff in to the memory.

SELECT table_name,inmemory,inmemory_priority,inmemory_distribute,inmemory_compression,inmemory_duplicate 
FROM user_tables WHERE table_name='TEST1';

TABLE_NAME           INMEMORY INMEMORY INMEMORY_DISTRI INMEMORY_COMPRESS INMEMORY_DUPL
-------------------- -------- -------- --------------- ----------------- -------------
TEST1                DISABLED

00:14:42 SQL> select pool, alloc_bytes, used_bytes, populate_status from v$inmemory_area;

POOL                       ALLOC_BYTES USED_BYTES POPULATE_STATUS
-------------------------- ----------- ---------- --------------------------
1MB POOL                     200278016          0 DONE                         >>>>>>> 1MB pool is used to store the column formatted data.
64KB POOL                     50331648          0 DONE                         >>>>>>> 64 KB pool is used to store the metadata about its residents.

Now i am moving the discussed table to in memory.

exec DBMS_INMEMORY.POPULATE(schema_name => dixit, table_name => 'TEST1');
or Using

00:15:01 SQL> alter table test1 inmemory;

Table altered.


00:17:59 SQL> SELECT table_name,inmemory,inmemory_priority,inmemory_distribute,inmemory_compression,inmemory_duplicate
FROM user_tables WHERE table_name='TEST1';

TABLE_NAME           INMEMORY INMEMORY INMEMORY_DISTRI INMEMORY_COMPRESS INMEMORY_DUPL
-------------------- -------- -------- --------------- ----------------- -------------
TEST1                ENABLED  NONE     AUTO            FOR QUERY LOW     NO DUPLICATE




00:20:17 SQL> select pool, alloc_bytes, used_bytes, populate_status from v$inmemory_area;

POOL                       ALLOC_BYTES USED_BYTES POPULATE_STATUS
-------------------------- ----------- ---------- --------------------------
1MB POOL                     200278016          0 DONE
64KB POOL                     50331648          0 DONE


00:18:25 SQL> select owner, segment_name, populate_status from v$im_segments;

no rows selected

Above results shows that no segment is added to the pool. Now lets try to query the table and see the results again.

00:20:26 SQL> select count(*) from test1;

  COUNT(*)
----------
   1000000



00:20:48 SQL> select pool, alloc_bytes, used_bytes, populate_status from v$inmemory_area;

POOL                       ALLOC_BYTES USED_BYTES POPULATE_STATUS
-------------------------- ----------- ---------- --------------------------
1MB POOL                     200278016          0 POPULATING
64KB POOL                     50331648          0 POPULATING


Okay now it's doing something as status from DONE has changed to POLULATING. Lets repeat the same command again.

POOL                       ALLOC_BYTES USED_BYTES POPULATE_STATUS
-------------------------- ----------- ---------- --------------------------
1MB POOL                     200278016   49283072 DONE
64KB POOL                     50331648     458752 DONE

Okay its there in the pool now! Lets check the in memory area, what we have inside it.

00:23:56 SQL> select owner, segment_name, populate_status from v$im_segments;

OWNER      SEGMENT_NAME         POPULATE_STAT
---------- -------------------- -------------
DIXIT      TEST1                COMPLETED

Lets conform if table is fully populated we have to look at v$im_segments_detail which compares the number of blocks in In-Memory and in the TEST1 table.

SELECT m.inst_id, 
       m.blocksinmem, 
       m.datablocks 
FROM   v$im_segments_detail m, 
       user_objects o 
WHERE  m.dataobj = o.object_id 
AND    o.object_name = 'TEST1';

   INST_ID BLOCKSINMEM DATABLOCKS
---------- ----------- ----------
         1        6922       6922

Lets see if there is any changes captured in Execution plan of the query.

SQL> explain plan for select count(*) from test1;

Explained.

SQL> @xplan

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------
Plan hash value: 3896847026

-----------------------------------------------------------------------------
| Id  | Operation                   | Name  | Rows  | Cost (%CPU)| Time     |
-----------------------------------------------------------------------------
|   0 | SELECT STATEMENT            |       |     1 |    71   (2)| 00:00:01 |
|   1 |  SORT AGGREGATE             |       |     1 |            |          |
|   2 |   TABLE ACCESS INMEMORY FULL| TEST1 |  1000K|    71   (2)| 00:00:01 |
-----------------------------------------------------------------------------

9 rows selected.

Now we have a new operation type added to the plan ‘TABLE ACCESS INMEMORY FULL’. It says that the object was accessed directly from the memory.

You can compare the response times with/without the In-Memory Store by simply disabling In-Memory functionality in your session:

ALTER SESSION SET inmemory_query = ENABLE;
ALTER SESSION SET inmemory_query = DISABLE;

You can anytime remove the objects from the pool. Other options to use with In memory.

SQL> ALTER TABLESPACE example INMEMORY;

Enabling the In-Memory attribute on the TEST1 table but excluding the “dest_id” column

SQL> ALTER TABLE TEST1 INMEMORY NO INMEMORY(dest_id);

Enabling the in memory option for table TEST1 and setting the priority to CRITICAL. With this option set on object (Other options are HIGH, MEDIUM, LOW), it will be populated immediately after the database is opened.

SQL> ALTER TABLE TEST1 INMEMORY PRIORITY CRITICAL;

There are few other options like Compression (Objects compressed during population) , Joins, Scans are also there and will be covered in a separate post or you can read the official documentation on in memory to understand them.

But i would like to discuss little about how the In-Memory option works in RAC environment.

In case of user querying the database in-memory in RAC, serial queries will only access a fraction of the data from its own node. IMCUs or IM Compression Units are not traveled using interconnect or using cache fusion. It fetches the remainder data from the disk itself.
Parallel execution helps as it starts multiple processes and ensures that at least one parallel server slave is allocated for each RAC instance. For that we have to set the parallel_degree_policy or Auto DOP to AUTO which makes the query coordinator ICMU aware and it (QC) automatically starts parallel server processes on correct server.

Now after all the discussion time to share my final words on the subject.
As we know that caching or pooling are the concepts there with Oracle RDBMS or with any of the databases from a very long time; then – why this in memory now ?
We already have few of the areas like KEEP/RECYCLE pools, Result cache etc. to keep cache data or results.
So my answer is that the Oracle In-Memory column store enables objects to load in memory in compressed columnar format which makes the scans to perform better that on-disk reads and leads to performance boosts. Use it when there is a SQL related performance issue and you’ve tried all tuning methods and each one of them are failed to fix the problem and yes If the In-Memory performance benefit can outperform the additional Oracle license costs.

Hope It Helps
Prashant Dixit