The innodb_flush_method is one of those variables that requires some thought. It is a factor in I/O responsiveness, while at the same time a contributor to durability (the D in Acid).
Some of the generalizations and misconceptions I’ve had about innodb_flush_method in the past:
- always use O_DIRECT – this is still recommended
- I thought when they meant “logs”, they meant all the logs (error, slow, general, audit) – this is untrue, they only mean the redo logs.
- double write buffer was the innodb buffer, redo log + page cache – this is also untrue. It is where write pages are stored prior to being flushed from the buffer pool. They are in fact within the ibdata file.
Primary Status: innodb_data_fsyncs
ACID component: durability
Note: Oracle MySQL and Percona MySQL now have an innodb_dedicated_server variable that will autotune innodb_buffer_pool_size, innodb_log_file_size and innodb_flush_method at startup. I’ve never tried it, but hope to test it out in the future. This variable will try to set O_DIRECT_NO_FSYNC if it can, otherwise fsync.