ZFS Shows Up in New Leopard Build

Udo Schmitz writes "As a follow-up to rumours from May this year, World of Apple has a screenshot showing Sun's Zettabyte File System in "the most recent Build of Mac OS X 10.5 Leopard". Though I still wonder: If it is not meant to replace HFS+, could there be any other reasons to support ZFS?"

Comments

Exciting!

[statusbar]

Now that Vista is finalized, expect Apple to show more and more of the 'secret' features of leopard!

--jeffk++

Reasons to support? Servers

[ShyGuy91284]

I will be soon converting my Linux server to Solaris just for ZFS. Although ZFS may not terribly useful on a normal desktop, on a server, it's very powerful.... The idea of parity data actually being used actively to ensure data isn't corrupted is brilliant imho. So is the idea of on-the-fly recovery (I remember a video of some guy writing 30 megs of junk to a partition using dd, ZFS detecting it, and repairing it). *ends rant since all this can be read up about online*

ZFS is overkill for a laptop - for now

[TheRaven64]

A few years ago, I sat down and worked out exactly what I thought a filesystem should do, and how I would implement it. At the time no filesystem came close. Then Sun released ZFS. Real documentation on it is hard to find (behind the marketing hype), but when I did track it down I discovered two things:

1. They had implemented everything I thought they should, and
2. That only accounted for about 40% of the features of ZFS.

Calling it the last word in filesystems might be hyperbole, but I expect ZFS to last a good 10-20 years, which is quite respectable for a filesystem, and I wouldn't be surprised if it lasted longer. Is it a replacement for HFS+? Not yet.

HFS+ is a very nice filesystem for single user systems with a single disk. It implements journalling, has reasonable performance, and has good metadata support. For the average users at the moment, the only real advantage of ZFS would be snapshots, and these are not too difficult to implement for other filesystems.

ZFS, however, is much better when you have multiple physical disks. At the moment, only the top-end Macs have more than one disk. This is likely to change in two ways:

1. Cheap flash,
2. Network storage

For a home user, ZFS could handle backups trivially by plugging in a large flash drive and adding it to the pool. I suspect this will be one mechanism Time Machine will use. Due to the way ZFS works, you can just mirror a part of the directory tree (e.g. /Users/aUser) onto the external disk. With a big external drive, you could mirror the entire disk onto it and also save snapshots (another Time Machine feature...). The same could be done with network storage. With the current price of hard drives, I wouldn't be surprised if .Mac started offering 10-20GB of storage space for remote backups using this mechanism (take a look at the NFS4 integration in ZFS to see how this could be done).

ZFS is not needed as a replacement for HFS+ in 2007, but it probably will be in 2008-9. ZFS is a 128-bit filesystem, which means it is designed to last for a long time. We will probably never need a 128-bit filesystem (unless we actually want to build hard drives the size of planets with single-atom sectors), but we will need a 65-bit filesystem once we get to around 10 Exabytes. This won't happen with single drives for a while, but it will with RAID arrays.

It's to support Time Machine

[FunWithHeadlines]

See this Ars Technica article where John Siracusa said back in August:

"For Mac geeks of a certain persuasion, the first mention of a soon-to-be-revealed feature of Leopard during the WWDC keynote set off a mental chain-reaction. That feature was Time Machine, and the name alone was enough to cause one particular phrase to hammer in the mind of many people, including me: "New file system in Leopard!" It was even a bingo square. In fact, it was my personal favorite bingo square, and the one that I most looked forward to marking.

But let's back up a bit. Why should the mere name "Time Machine" scream "new file system" to anyone? And why the excitement about a new file system in the first place? What's wrong with HFS+, Mac OS X's current file system? It's got journaling. It supports arbitrarily extensible metadata. It can even be case-sensitive to satisfy the Unix geeks. Does Mac OS X really need a new file system?

In a word, yes. HFS was a state-of-the-art personal computer file system when it was first released...twenty-one years ago. HFS+ is only eight years old, but it's built on many of the design decisions of HFS. Progress marches on. Today, there are new capabilities that the best modern file systems have, but that HFS+, even with all of its recent additions, does not. Here's a short list.

• Efficient storage and handling of very small files.
• Logical volume management through a pooled storage model.
• Improved data integrity using checksums on all data.
• Snapshots.

So it's about the snapshot ability of ZFS, and that's exactly what will be needed for Time Machine.

Re:copy-on-write

[TheRaven64]

No. Mmap lives above the filesystem layer. Unless you are doing mmap on the block device, in which case you should realise that not everyone works for oracle...

Mmap simple maps pages of a disk file into memory. If the disk file changes its physical location then the mapping is updated. When you call mmap, you give it a disk file, an offset, and an extent. It is up to the VFS layer to translate this into physical mappings. LFS has the same issues, and these were solved well over a decade ago.

If you invoke mmap with MAP_PRIVATE, this actually makes it easier; if someone else updates the file then you just keep the existing mapping.

Re:copy-on-write

[Midnight Thunder]

FTA:

Makes use of copy-on-write; rather than overwriting old data with new data, it writes new data to a new location and then overwrites the pointer to the old data

Wouldn't that pose a problem for mmap?

It may do, but like many things there are alternative approaches.

From working on embedded hardware with flash memory, this makes me wonder whether possible addition of ZFS is meant to be for flash storage? Let me explain: flash memory has a fairly limited write-count, relative to hard disks, so to compensate for this memory is written in a circular fashion, to ensure that a given sector is written the least often possible. In addition to this, from what I can tell, Apple's main sales point are low profile computers and portables. The latter would benefit from flash storage as means of extending battery life, even if it is for a certain elements, such as for the OS which is accesed far more frequently than anything else on disk. Given this I wouldn't be surprised to see flash memory in future models of Apple portables, using ZFS, while HFS+ is still used for the hard disks.

This is pure speculation, but I feel that it has a high probabilty of being near the mark.

...so how does one define "capacity" therein?

[cyclomedia]

I've been wandering about this and am insanely curious: if ZFS really does intelligently copy on write how far can it take it?

for starters, does the FS "know" that i've just clicked "Save As" in my word processor? what about copy and pasting a file back into the same directory to make a local copy? Also? is it just within variations on the same file? if i have a particular setup exe on my system but forget, and download it again to the desktop surely the FS has no initial way of knowing that they are one and the same, does some funky heuristic happen?

basically: does the OS's read/write/copy/delete functionality have to invoke copy-on-write via a FS API or is it built in for every single sector-sized chunk that gets stuffed into the FS?

the next question is the one in my subject: how therefore do you define "capacity"? if i've got a bunch of files that take up 700mb on a ZFS device and try to back up to a (Joliet) CD will i get a message telling me that the CD doesnt have room? i can imagine this scenario being unlikely with optimised binary data (jpegs and mpegs) but if i'm backing up a dev environment with autobackups (main.c,main.c.bak.001,main.c.bak.002,etc.) and manually created and dated directory tree "snapshots" (dev,dev_backup_2006-12-18,dev_backup_2006-12-01,e tc.) then this could probably happen quite easily.