A proposal to make disk IO much faster under SSD

[suiyemin sui]

Problem: CentOS, insert+commit operation is very slow. There are 3 types of files need to be synced to ensure data persistence: innodb data files, log files and binary log files.

In current implementation, for each commit the fsync or fdatasync will be invoked to ensure data persistence. That will cost 3~4 milliseconds. That's necessary for HDD. However, there is a much better way for SSD.

Proposal under SSD: use directIO for all those files without data sync. For volatile cache inside SSD, just disable it.(For industry level SSD, there will be a non-volatile cache inside SSD which will make SSD fast enough.) In this way, the commit operation will cost only about 200 microseconds.

I know it's not easy to support full directIO. But it will be fascinating fast.

Best Regards,
Yemin