Tables with Memoptimize for Read (fast lookup)

Hi,

I would like to share my initial thoughts about memoptimized for read (fast lookup) tables in this post. fast lookup tables came with 18c and they are here to provide faster access to a table if you are using primary key with a equality condition. this is basically single table hash cluster. instead of using primary key index, Oracle creates a hash index and searching on a hash index should be much more faster than a unique index.

first, create the test environment.

to use a fast lookup table you must set memoptimize_pool_size paramter to a number higher than 0. this is a static memory component in sga to store hash index of the table. this memory component is not resizable which means it won’t grow or become smaller with automatic shared memory management.

create a table for fast lookup:

segment creation must be immediate otherwise you will get an error! Your table must have a primary key as well, since whole structure is based on PK, it is logical to have one. I roughly inserted 146.000 rows into table.

EDIT: I also tried those tests with 14 million rows too but results are similar.

this is not enough! you must use dbms_memoptimize package to use this table as fast lookup table.

let’s see what is happening now:

I skipped first execution statistics! you will see a small consistent gets at the first execution but consecutive executions will be like above.

as you see there are new execution plan steps “TABLE ACCESS BY INDEX ROWID READ OPTIM” and “INDEX UNIQUE SCAN READ OPTIM”, so these steps tell us, an hash index is used to retrieve data like as in key-value pairs. there is no “consistent gets”. this is amazing. almost nothing has read for this query and that should make this query so much faster.  Is it? well, I couldn’t find it, not as much as I expected at least. Actually, I found different blog posts mentioning about fast lookup tables but non of them has made a performance test.

So, I want to test the speed but there is a problem. for now, there are many limitations on fast lookup tables. one of them is “you cannot use it via plsql”. this means I can’t create a simple plsql procedure and run a query for 100000 times and compare timings. to do a test, I wrote a small python code.

Warning: I am no python developer, I just now python to solve my small problems, this code that I share below probably has many bugs but I just use it for this test purposes.

here is the python code I use:

after changing code many times, I used this final code. this code creates 9 threads and every thread run “select * from tmp_optimize where id = :myid” query for 100.000 times and myid variable value is between 1 and 10000 in circular way in this loop. each threads print total execution time for 100.000 runs of the query. here is the result:

timing is changed between 16.0 and 16.1 seconds for this fast lookup tables. how about a normal table with normal primary key?

I used same python code just changed table name from “tmp_optimize” to “tmp_not_optimize” and result is like this:

it is between 17.1 and 17.2 seconds. there is %6 performance gain (give or take).

I am not sure whether %6 is a good gain or not because to use fast lookup table actively, we need to separate serious amount of memory area. Also, there are few limitations like, cannot be used with partitioned or compressed tables, cannot be used in plsql (not yet anyway). only equality conditions on primary keys can advantage of fast lookup.

During my tests I didn’t check wait events. I expect that less latch waits on fast lookup table since no consistent gets occur. Maybe there is something I missed and I will do more tests on these tables (I didn’t even test “memoptimize for write” (fast ingest) yet).

wish you all healthy days.

Leave a Reply

Your email address will not be published.