page alignment for large malloc requests?

[Florian Weimer]

* Andreas Dilger:

> No, because we can't know in advance whether all IO on the fd will
> be properly aligned and sized for O_DIRECT.  This would be decided
> on a per-write*() syscall basis, if it is large enough and aligned
> properly.  We saw good improvements for some IO benchmarks with
> automatic O_DIRECTifying large write calls, […]

I do not think this would result in an overall benefit.  Have you
tweaked the kernel to evcit the modified pages from the page cache after
large writes, to fake the impact of the change?  And run some typical
tasks with that, like a kernel build?

I expect that performance will not be great because most written data is
read again after a fairly short time.

Thanks,
Florian