I haven't found many fresh benchmarks regarding the issue. The most relevant link I've seen is [1], which shows better performance of O_DIRECT in SAN, there is also detailed comment written by engineer from Oracle about many factors affecting the results. IMHO it may be better to just expose innodb_flush_method in puppet to let deployment engineers decide and explicitly set it to the default MySQL value – fdatasync.
I haven't found many fresh benchmarks regarding the issue. The most relevant link I've seen is [1], which shows better performance of O_DIRECT in SAN, there is also detailed comment written by engineer from Oracle about many factors affecting the results. IMHO it may be better to just expose innodb_flush_method in puppet to let deployment engineers decide and explicitly set it to the default MySQL value – fdatasync.
[1] http:// themattreid. com/wordpress/ 2012/01/ 06/san- vs-local- disk-innodb_ flush_method- performance- benchmarks/