Slower performance with ext4
Chan Chung Hang Christopher
christopher.chan at bradbury.edu.hk
Mon Nov 2 13:37:59 UTC 2009
Mark Kirkwood wrote:
> Christopher Chan wrote:
>> Yawn. Been there and done that. Without bbu cached hardware raid. Just
>> plain Linux software raid. XFS = pray for no power loss and ext3
>> data=journal = sleep well at night (except for spammers getting
>> through the developers' webmail system).
>> You are using hardware raid + bbu and you have no need to delve deep
>> into how the filesystems work. If you do not want to take even the
>> standard explanations for ext3's (which are repeated for ext4) different
>> journaling modes then that is just too bad. Just stop propagating the
>> myth that fsync = return after data has been written to the filesytem.
>> If that was the case, there would not be large differences in filesystem
> Double yawn - of course there is a *performance* difference - different
> filesystems do writes different ways (extent vs not for instance),
> however reliability is determined by the interaction with the device
> layers write cache (amongst other things - but that is the main effect
> here). Not matter how clever the filesystem - if the underlying hardware
> does not actually do the writes as requested (queue cheap sata as I've
> mentioned), then all bets are off. Hence the market for battery backed
> raid controllers for sata drives in particular (and note that these are
> worth using particularly 3Ware and Areca).
Heh. So do you care then to explain why there is a performance
difference within the SAME filesystem due to different journaling modes
chosen then? fsbench was first written to see the differences between
the various modes of ext3 and also to compare other filesystems with or
without external journals.
> So to reiterate - guaranteeing writes to filesystem is good - but not
> good enough if the underlying device does not honor the software
> request. This is the guts of most workstation corruption problems,
> regardless of fs type.
> For instance, I have experienced power interruption data loss on my
> workstation (ext3 filesystem + cheap sata drive) - and this is expected
> from this type of hardware.
Too bad that fsync in Linux does not flush write-caches whether
ide/sata/scsi if enabled. It does not matter whether the disk honours
the software request because there is no such request. Known problem
since 2001 and still present until at least 2009 and the fsync man page
does not indicate any change. The kind of disk has nothing to do with it.
More information about the ubuntu-users