> > Keep in mind that we support platforms without O_DSYNC. I am not
> > sure whether there are any that don't have O_SYNC either, but I am
> > fairly sure that we measured O_SYNC to be slower than fsync()s on
> > some platforms.
This measurement is quite understandable, since the current software
does 8k writes, and the OS only has a chance to write bigger blocks in the
write+fsync case. In the O_SYNC case you need to group bigger blocks yourself.
(bigger blocks are essential for max IO)
I am still convinced, that writing bigger blocks would allow the fastest
solution. But reading the recent posts the solution might only be to change
the current "loop foreach dirty 8k WAL buffer write 8k" to one or two large
write calls.
Andreas