Re: [gentoo-user] ZFS

[Pandu Poluan]

On Tue, Sep 17, 2013 at 2:20 PM, Grant <emailgrant@×××××.com> wrote:
> I'm convinced I need 3-disk RAID1 so I can lose 2 drives and keep
> running.  I'd also like to stripe for performance, resulting in
> RAID10.  It sounds like most hardware controllers do not support
> 6-disk RAID10 so ZFS looks very interesting.
>
> Can I operate ZFS RAID without a hardware RAID controller?
>

Yes. In fact, that's ZFS' preferred mode of operation (i.e., it
handles all redundancy by itself).

> From a RAID perspective only, is ZFS a better choice than conventional
> software RAID?
>

Yes.

ZFS checksummed all blocks during writes, and verifies those checksums
during read.

It is possible to have 2 bits flipped at the same time among 2 hard
disks. In such case, the RAID controller will never see the bitflips.
But ZFS will see it.

> ZFS seems to have many excellent features and I'd like to ease into
> them slowly (like an old man into a nice warm bath).  Does ZFS allow
> you to set up additional features later (e.g. snapshots, encryption,
> deduplication, compression) or is some forethought required when first
> making the filesystem?
>

Snapshots is built-in from the beginning. All you have to do is create
one when you want it.

Deduplication can be turned on and off at will -- but be warned: You
need HUGE amount of RAM.

Compression can be turned on and off at will. Previously-compressed
data won't become uncompressed unless you modify them.

> It looks like there are comprehensive ZFS Gentoo docs
> (http://wiki.gentoo.org/wiki/ZFS) but can anyone tell me from the real
> world about how much extra difficulty/complexity is added to
> installation and ongoing administration when choosing ZFS over ext4?
>

Very very minimal. So minimal, in fact, that if you don't plan to use
ZFS as a root filesystem, it's laughably simple. You don't even have
to edit /etc/fstab

> Performance doesn't seem to be one of ZFS's strong points.  Is it
> considered suitable for a high-performance server?
>
> http://www.phoronix.com/scan.php?page=news_item&px=MTM1NTA
>

Several points:

1. The added steps of checksumming (and verifying the checksums)
*will* give a performance penalty.

2. When comparing performance of 1 (one) drive, of course ZFS will
lose. But when you build a ZFS pool out of 3 pairs of mirrored drives,
throughput will increase significantly as ZFS has the ability to do
'load-balancing' among mirror-pairs (or, in ZFS parlance, "mirrored
vdevs")

Go directly to this post:
http://phoronix.com/forums/showthread.php?79922-Benchmarks-Of-The-New-ZFS-On-Linux-EXT4-Wins&p=326838#post326838

Notice how ZFS won against ext4 in 8 scenarios out of 9. (The only
scenario where ZFS lost is in the single-client RAID-1 scenario)

> Besides performance, are there any drawbacks to ZFS compared to ext4?
>

1. You need a huge amount of RAM to let ZFS do its magic. But RAM is
cheap nowadays. Data... possibly priceless.

2. Be careful when using ZFS on a server on which processes rapidly
spawn and terminate. ZFS doesn't like memory fragmentation.

For point #2, I can give you a real-life example:

My mail server, for some reasons, choke if too many TLS errors happen.
So, I placed "Perdition" in to capture all POP3 connections and
'un-TLS' them. Perdition spawns a new process for *every* connection.
My mail server has 2000 users, I regularly see more than 100 Perdition
child processes. Many very ephemeral (i.e., existing for less than 5
seconds). The RAM is undoubtedly *extremely* fragmented. ZFS cries
murder when it cannot allocate a contiguous SLAB of memory to increase
its ARC Cache.

OTOH, on another very busy server (mail archiving server using
MailArchiva, handling 2000+ emails per hour), ZFS run flawlessly. No
incident _at_all_. Undoubtedly because MailArchiva use one single huge
process (Java-based) to handle all transactions, so no RAM
fragmentation here.


Rgds,
-- 
FdS Pandu E Poluan
~ IT Optimizer ~

 • LOPSA Member #15248
 • Blog : http://pepoluan.tumblr.com
 • Linked-In : http://id.linkedin.com/in/pepoluan