mysql setting variable innodb_flush_method to O_DSYNC or O_DIRECT

Question

In my configuration innodb_flush_method=O_DSYNC from O-DIRECT reduces about 75% the iowait, and accordingly this the load. Should I set another variables besides innodb_flush_method to reduce more the iowait?

My configuration file is:

[mysqld]
innodb_file_per_table=1
query_cache_size=128M
thread_cache_size=64
key_buffer_size=32M
max_allowed_packet=16M
table_cache=1024
table_definition_cache=8192
wait_timeout=20
max_user_connections=25
innodb_flush_method=O_DSYNC
open_files_limit=16384

myisam_sort_buffer_size=2M

collation_server=utf8_unicode_ci
character_set_server=utf8

tmp_table_size = 384M
max_heap_table_size = 384M
innodb_buffer_pool_size=64M
innodb_thread_concurrency=8
max_connections=125

I have a database with 100 Innodb tables, 3 of them has about 25000 records, the others has no significant records. The average queries in peak time is about 160, the majority is SELECT

Answer 1

innodb_buffer_pool_size

Major problem is innodb_buffer_pool_size is too small. Recommandation is set to 50~75% of main memory.

innodb_buffer_pool_size=64M

I strongly recommand that you should increase it's value.

Generally speaking, O_DIRECT is little bit fast because InnoDB Buffer Pool caches Data+Index, So with O_DIRECT disabled File System Page Cache is faster. MySQL Manaual says ( http://dev.mysql.com/doc/refman/5.1/en/innodb-parameters.html#sysvar_innodb_flush_method )

Depending on hardware configuration, setting innodb_flush_method to O_DIRECT can either have either a positive or negative effect on performance. Benchmark your particular configuration to decide which setting to use.

But in my experience, there was no significant difference between O_DIRECT and O_DSYNC. Both SSD and HDD are test.

Anyway you should increase innodb_buffer_pool_size .

Calculating innodb buffer pool hit ratio

mysql> SHOW GLOBAL STATUS LIKE '%innodb%';
+---------------------------------------+-------------+
| Variable_name                         | Value       |
+---------------------------------------+-------------+
.....
.....
| Innodb_buffer_pool_read_requests      | 11054273949 |
| Innodb_buffer_pool_reads              | 135237      |
| Innodb_buffer_pool_wait_free          | 0           |
....

innodb buffer pool hit ratio = ((Innodb_buffer_pool_read_requests) / (Innodb_buffer_pool_read_requests + Innodb_buffer_pool_reads)) * 100

Eg above examples,

hit ratio = (11054273949  / (11054273949  + 135237)) * 100 = 99.99%

I think this value is too small in your case.

query_cache_size

"the majority is SELECT"

If most queries are SELECT and update query is rare, I think increasing query_cache_size is very helpful for you.

Could you post your query cache status as follows?

mysql> show global status like 'Qc%';
+-------------------------+------------+
| Variable_name           | Value      |
+-------------------------+------------+
| Qcache_free_blocks      | 13         |
| Qcache_free_memory      | 1073403104 |
| Qcache_hits             | 217949     |
| Qcache_inserts          | 337009     |
| Qcache_lowmem_prunes    | 0          |
| Qcache_not_cached       | 2122598    |
| Qcache_queries_in_cache | 68         |
| Qcache_total_blocks     | 167        |
+-------------------------+------------+

mysql> show global status like 'com_select%';
+---------------+---------+
| Variable_name | Value   |
+---------------+---------+
| Com_select    | 3292531 |
+---------------+---------+
1 row in set (0.00 sec)

Calculating innodb buffer pool hit ratio

 query cache hit ratio = ((Qcache_hits) / (Qcache_hits + Com_select)) * 100

first, figure out your query cache hit ratio.