>we'd know how to get the best performance when writing
>absolutely synchronously ie.  the data had to actually
>end up on a disk platter and not just get parked in
>some cache, like in the buffer cache or in the drive
>controller's cache.

I should have mentioned why we instrumented rewrite
versus append performance, since it's apropos the
current metadata overhead discussion.  When an existing
file (particularly a large file) is rewritten the
filesystem code has the opportunity (though not the
obligation) to reuse the existing block allocations;
this is almost always a win as it requires far less
manipulation of the metadata to write the same amount
of "real" data.  If you instead first delete and then
recreate the file from nothing, you typically have to
fondle the metadata a whole lot more, which takes time.
_______________________________________________
gnhlug-discuss mailing list
[EMAIL PROTECTED]
http://mail.gnhlug.org/mailman/listinfo/gnhlug-discuss

Reply via email to