> Our experience with various Unix flavors teaches us that data-only sync > does not have a performance advantage over a full sync.
At first glance, this seems surprising. What is the typical split between fsync'ed writes that are changing the file size versus such that are not changing the file size? On Linux/ext4, the former should not see a performance improvement but the latter should. Eg., there are synthetic benchmark results at https://crbug.com/469071 which show 2x improvment of fdatasync over fsync for a system that is under heavy I/O load. Best regards, Thiemo Nagel