[ale] ISCSI array on virtual machine

[Jim Kinney]

If you need de-dup, ZFS is the only choice and be ready to throw a lot
of RAM into the server so it can do it's job. I was looking at dedupe
on 80TB and the RAM hit was 250GB.
XFS vs EXT4.
XFS is the better choice.
XFS does everything EXT4 does except shrink. It was designed for (then
very) large files (video) and works quite well with smaller files. It's
as fast as EXT4 but will handle larger files and many, many more of
them. I want to say exabytes but not certain. Petabytes are OK
filesystem sizes with XFS right now. I have no experience with a
filesystem of that size but I expect there to be some level of metadata
performance hit.
If there's the slightest chance of a need to shrink a partition (You
_are_ using LVM, right?) then XFS will bite you and require relocation,
tear down, rebuild, relocation. Not a fun process.
A while back, an install onto a 24 TB RAID6 array refused to budge
using EXT4. While EXT4 is supposed to address that kind of size, it had
bugs and unimplemented plans for expansion features that were blockers.
I used XFS instead and never looked back. XFS has a very complete
toolset for maintenance/repair needs.  
On Wed, 2016-04-27 at 13:54 -0500, Todor Fassl wrote:
> I need to setup a new file server on a virtual machine with an
> attached 
> ISCSI array. Two things I am obsessing over -- 1. Which file system
> to 
> use and 2. Partitioning scheme.
> 
> The ISCSI array is attached to a ubuntu 16.04 virtual machine. To
> tell 
> you the truth, I don't even know how that is done. I do not manage
> the 
> VMware cluster.  In fact, I think the Dell technitian actually ddid
> that 
> for us. It looks like a normal 8T hard drive on /dev/sdb to the
> virtual 
> machine. The ISCSI array is configured for RAID6 so from what I 
> understand, all I have to do is choose a file system appropriate for
> my 
> end user's needs. Even though the array looks like a single hard
> drive, 
> I don't have to worry about software RAID or anyhthing like that.
> 
> Googling shows me no clear advantage to ext4, xfs, or zfs. I haven't 
> been able to find a page that says any one of those is an obvious
> choice 
> in my situation. I have about 150 end-users with nfs mounted home 
> directories. We also have a handful of people using Windows so the
> file 
> server will have samba installed. It's a pretty good mix of large
> files 
> and small files since different users are doing drastically
> different 
> things. There are users who never do anything but read email and
> browse 
> the web and others doing fluid dynamic simulations on small
> supercomputers.
> 
> Secondthing I've been going back and forth on in my own mind is
> whether 
> to do away with seperate partitions for faculty, staff, and grad 
> students. My co-worker says that's probably an artifact of the days
> when 
> partition sizes were limited. That was before my time here. The last
> 2 
> times we rebuilt our file server, we just maintained the
> partitioning 
> scheme and just made the sizes  times larger. But sometimes the
> faculty 
> partition got filled up while there was still plenty of space left
> on 
> the grad partition. Or it might be the other way around. If we
> munged 
> them all together, that wouldn't happen. The only downside I see to 
> doing that is that if the faculty partition gets hosed, the grad 
> partition wouldn't be effected. But that seems like a pretty
> arbitrary 
> choice. We could just assign users randomly to one partition or
> another. 
> When you're setting up a NAS for use by a lot of users, is it
> considered 
> best practice to split it up to limit the damage from a messed up
> file 
> system? I mean, hopefully, that never happens anyway, right?
> 
> Right now, I've got it configured as one gigantic 8T ext4 partition.
> But 
> we won't be going live with it until the end of May so I have plenty
> of 
> time to completely rebuild it.
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.ale.org/pipermail/ale/attachments/20160427/983cb5e6/attachment.html>