On Tue, 8 May 2007, Greg Smith wrote:
On Tue, 8 May 2007, Luke Lonergan wrote:
From discussions with the developers, the biggest issue is a technical
one: the Linux VFS layer makes the [ZFS] port difficult.
Difficult on two levels. First you'd have to figure out how to make it work
at all; then you'd have to reshape it into a form that it would be acceptable
to the Linux kernel developers, who haven't seemed real keen on the idea so
far.
given that RAID, snapshots, etc are already in the linux kernel, I suspect
that what will need to happen is for the filesystem to be ported without
those features and then the userspace tools (that manipulate the volumes )
be ported to use the things already in the kernel.
The standard article I'm you've already seen this week on this topic is Jeff
Bonwick's at http://blogs.sun.com/bonwick/entry/rampant_layering_violation
yep, that sounds like what I've been hearing.
what the ZFS (and reiserfs4) folks haven't been wanting to hear from the
linux kernel devs is that they are interested in having all these neat
features available for use with all filesystems (and the linux kernel has
a _lot_ of filesystems available), with solaris you basicly have UFS and
ZFS so it's not as big a deal.
What really bugged me was his earlier article linked to there where he talks
about how ZFS eliminates the need for hardware RAID controllers:
http://blogs.sun.com/bonwick/entry/raid_z
While there may be merit to that idea for some applications, like situations
where you have a pig of a RAID5 volume, that's just hype for database writes.
"We issue the SYNCHRONIZE CACHE command to the disks after pushing all data
in a transaction group"--see, that would be the part the hardware controller
is needed to accelerate. If you really care about whether your data hit
disk, there is no way to break the RPM barrier without hardware support. The
fact that he misunderstands such a fundamental point makes me wonder what
other gigantic mistakes might be buried in his analysis.
I've seen similar comments from some of the linux kernel devs, they've
used low-end raid controllers with small processors on them and think that
a second core/socket in the main system to run software raid on is better.
David Lang
