softupdates performance...

[Matthew Dillon]

    FreeBSD-5 has a lot overhead, especially when it comes to the buffer
    manipulation that softupdates does.  DFly should be nearly the same as
    4.x in regards to FS performance.

    There isn't much I can do about softupdates but the (slowly progressing)
    namecache work will eventually allow us to release the exclusive lock
    on the directory vnode during directory searches and this will bring
    up our lots-of-little-file benchmark numbers considerably.

    Another issue that slows down filesystem operations is the busy-page
    lockout that occurs when the system is writing data to disk and some
    other operation wants to modify the page undergoing I/O.  That is
    'on the table' as well.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>


:I was interested to read this posting from Bruce Evans on freebsd-current:
:
:http://docs.FreeBSD.org/cgi/mid.cgi?20040527120819.B8434
:   soft
:   updates uses extra CPU cycles to try to speed up i/o to real drives (and
:   lately it doesn't seem to be very successful in doing the latter -- here
:   it is now about the same speed as normal mounts for copying /usr/src but
:   was 1.5 times faster a few years ago; async mounts are still 2.5 times
:   faster).
:
:I haven't tested softupdates vs non-softupdates on my DFly slice, but
:my subjective impression is that it is indeed much slower than it used
:to be: any disk IO takes ages, and slows down everything else too.
:It certainly "seems" like it was much faster on FreeBSD 4.x a year ago
:on my old (and much slower) laptop.  Linux ext3 seems much faster too.
:
:Any ideas?
:
:Rahul