Subject: Re: File system performance on i386
To: Christoph Hellwig <hch@lst.de>
From: Jaromír <jdolecek@netbsd.org>
List: tech-perform
Date: 02/23/2001 10:13:12
Christoph Hellwig wrote:
> More or less. Actually it avoids ffs complexity where it is not worth the
> effort with modern hardware/OS-Design (fragments, knowledge of physical
> media layout)

IIRC ext2fs doesn't have fragments, but some quite similar concept.
Fragments are quite useful to safe disk space, with any combination
of hardware and OS.

> actually asynch metadata increaese the reliablity

Good joke :)

> - unlike synch metadata the io is done in (mosty) the same
> order - of metadata is actually recoverable by a good fsck anyway.
> Basically what Linux does is softdep for the poor.

The problem is that since the metadata are written asynchronously
(read as "at some undefined time in future" and "at undefined
order"), it may not have hit the disk yet after the change or may
only hit the disk partially. So you CAN'T recover it. If you write
metadata synchronously (read "immediatelly"), you can be sure all
the changed metadata HIT the disk, so potential crash or anything
is not such disastrous. softdep maintains best of both words - it
postpones writing the metadata similar way as async does, but writes
all related metadata together, so that it garantees the state on
disk is consitent.

Try sometimes some bigger disk activity (say, some untarring, deleting)
on Linux and BSD. Hit the power switch in the middle of operation.
Then compare the results on Linux and BSD after the disks are
fscked :)

If you really really think async metadata is the way to go, you can
mount async if you prefer to. IIRC there is no option to write
only metadata sync on ext2fs under Linux though.

Jaromir 
-- 
Jaromir Dolecek <jdolecek@NetBSD.org>      http://www.ics.muni.cz/~dolecek/
@@@@  Wanna a real operating system ? Go and get NetBSD, dammit!  @@@@