After examining the benchmark design - I see we are probably not being
helped by the repeated insertion of keys all of form 'userxxxxxxx'
leading to some page splitting.
However your index rebuild gets you from 5 to 3 GB - does that really
help performance significantly?
regards
Mark
On 11/08/16 16:08, Kisung Kim wrote:
> Thank you for your information.
> Here is the result:
>
> After insertions:
>
> ycsb=# select * from pgstatindex('usertable_pkey');
> version | tree_level | index_size | root_block_no | internal_pages |
> leaf_pages | empty_pages | deleted_pages | avg_leaf_density |
> leaf_fragmentation
>
---------+------------+------------+---------------+----------------+------------+-------------+---------------+------------------+--------------------
> 2 | 3 | 5488721920 | 44337 | 4464 |
> 665545 | 0 | 0 | 52 | 11
> (1 row)
>
> After rebuild:
>
>
> ycsb=# select * from pgstatindex('usertable_pkey');
> version | tree_level | index_size | root_block_no | internal_pages |
> leaf_pages | empty_pages | deleted_pages | avg_leaf_density |
> leaf_fragmentation
>
---------+------------+------------+---------------+----------------+------------+-------------+---------------+------------------+--------------------
> 2 | 3 | 3154296832 | 41827 | 1899 |
> 383146 | 0 | 0 | 90.08 |
> 0
>
>
> It seems like that rebuild has an effect to reduce the number of
> internal and leaf_pages and make more dense leaf pages.
>
>
>