[subego]

Thanks. Crystal clear now.

[via Windows Weekly]

[OreoCookie]

ReFS is largely a reaction to ZFS (and btrfs which was also started as a GPL-compatible »clone«*). IMHO this is a good thing for Microsoft, and let's hope Sinofski can keep up his record of delivering features. I just noticed how old NTFS is which is a lot more modern than HFS+. Now it's Cupertino's turn to deliver a new file system … 

* I use the word clone very loosely here: it means to implement the same basic paradigms of ZFS (it also violates the traditional *nix layers as it combines filesystem, volume manager and RAID controller, it has copy on write, checksumming, etc.).

[subego]

So what you're saying is Apple's new file system shouldn't have a wave thingy?

[OreoCookie]

No, I don't care if it's wavy, curly or zig-zag, as long as Apple can get its act together and replace HFS+.

[subego]

Dig your own grave then. The graphic explicitly shows flat is superior to wavy.

[Big Mac]

If you look at Gruber's review of Lion and his discussion of the current state of the filesystem, what he discusses about what Apple does in Lion in implementing full disk Filevault sounds a lot like what Microsoft is describing in that slide. They're both pursuing a method to modernize their filesystems without completely throwing them out (to maintain backward compatibility chiefly, I suppose). The top layer of the filesystem will remain traditional, but below it sits a modern filesystem.

[P]

Originally Posted by subego 

Dig your own grave then. The graphic explicitly shows flat is superior to wavy.

The wavy vs. flat is a subtle point that I believe they made in the presentation - I have seen it mentioned in passing in reports, anyway. Generally an OS should be organized hierarchically in that there is a kernel and then libraries in layers, with no library relying on a function from a library at a higher level. Apple does this by clearly separating the UI libraries in Cocoa Appkit and Carbon from the underlying libraries in Foundation & CoreFoundation. These in turn rely on lower level libraries in the UNIX userland, etc. Any Linux desktop works the same way: Gnome/KDE rely on X11 which relies on lower level libraries. You cannot have an X11 library importing a function from Gnome any more than you can have a CoreFoundation library importing a Cocoa function.

Windows does not have this enforced hierarchy. Any lib can import a function from any other, and in practice they do. There is e.g. no way to run a command line only Windows without loading a lot of libraries that were made for the GUI. This means that isolating bugs is very hard, and replacing any part of the system is close to impossible. It was one of the reasons for the Vista disaster. The same thing can be blamed for the problems Apple had with Copland.

MS is (now) well aware that this is a problem, and they have a project called Miniwin that aims to untangle the spaghetti. The image is meant to say that ReFS is untangled. My guess is that NTFS relied on "higher" libraries for things like Unicode text, compression and encryption, and that ReFS solves that by reimplementing such functions at the appropriate level of the stack.

[besson3c]

I expect that Apple is working on a ZFS/BTRFS clone of their own, and like I have babbled about before, this will be a complete game changer for every Mac user - it will change so much including potentially how Apple designs hardware.

However, I can foresee this happening incrementally, the first big and stressful one getting users off of HFS+ and onto the new file system without making anything explode. I'd expect that the initial release won't include as many features as ZFS has today, but that over time Apple will start moving some of their existing features to operate at the filesystem level. So many of the features that are in OS X today would make sense to work at the filesystem, but I suspect that Apple is deathly afraid to make too many changes to HFS+ at this point.

Features that I think will be eventually reworked to involve the new file system:

- Time Machine
- Versions
- Auto save (snapshots instead of complete file copies, possibly to iCloud)
- Spotlight
- Obviously Disk Utility, Boot Camp, Lion's recovery partition, other stuff that deals with changing partition maps
- Filevault
- File sharing
- Finder metadata


Other improvements I can foresee:

- Auto file compression setting for faster disk access
- Using iCloud for storing checksum data copies of files for self heal so that single disk systems never have to deal with data corruption (no more fsck, Disk Warrior, etc.) This might take some time before it is cost effective for Apple and so that this works well without requiring a whole lot of bandwidth, but I think we'll eventually get there. Maybe OS X will queue up parity data and send it to iCloud during idle computer/network time or something
- Obviously the ability to create RAID arrays, but this time including RAID levels 5, 6, 10, etc.
- Apple providing separate SSD disks for write buffering in new hardware
- The ability to create pools of data that live primarily on iCloud, making a sort of Github like system for working on collaborative projects with users (rather than having to constantly upload complete copies of files like iCloud must do now)
- The ability to downgrade and/or switch between software versions of things without having to do a complete reinstall
- pool quotas
- deduplication
- file ACLs for users/groups
- different volume block sizes for different directories/pools (which might have an impact on performance in addition to space usage)
- more self-awareness and OS integration, such as notifying you when a disk is failing

I'm sure if I thought some more I could think of some more things, but I hope you can see why I feel the way I do about the magnitude of potential here, in the long run. Apple will be laughing once they get people off of HFS+ onto version one of this new file system, cause it will no doubt be far easier to roll out file system improvements than it is today with HFS+. Today ZFS/zpool updates can be applied by the user at their leisure, and they are pushed out without having to update the entire OS.

[subego]

Originally Posted by P 

The wavy vs. flat is a subtle point that I believe they made in the presentation - I have seen it mentioned in passing in reports, anyway. Generally an OS should be organized hierarchically in that there is a kernel and then libraries in layers, with no library relying on a function from a library at a higher level. Apple does this by clearly separating the UI libraries in Cocoa Appkit and Carbon from the underlying libraries in Foundation & CoreFoundation. These in turn rely on lower level libraries in the UNIX userland, etc. Any Linux desktop works the same way: Gnome/KDE rely on X11 which relies on lower level libraries. You cannot have an X11 library importing a function from Gnome any more than you can have a CoreFoundation library importing a Cocoa function.

Windows does not have this enforced hierarchy. Any lib can import a function from any other, and in practice they do. There is e.g. no way to run a command line only Windows without loading a lot of libraries that were made for the GUI. This means that isolating bugs is very hard, and replacing any part of the system is close to impossible. It was one of the reasons for the Vista disaster. The same thing can be blamed for the problems Apple had with Copland.

MS is (now) well aware that this is a problem, and they have a project called Miniwin that aims to untangle the spaghetti. The image is meant to say that ReFS is untangled. My guess is that NTFS relied on "higher" libraries for things like Unicode text, compression and encryption, and that ReFS solves that by reimplementing such functions at the appropriate level of the stack.

That's exactly what I said.

[Big Mac]

Originally Posted by P 

The wavy vs. flat is a subtle point that I believe they made in the presentation - I have seen it mentioned in passing in reports, anyway. Generally an OS should be organized hierarchically in that there is a kernel and then libraries in layers, with no library relying on a function from a library at a higher level. Apple does this by clearly separating the UI libraries in Cocoa Appkit and Carbon from the underlying libraries in Foundation & CoreFoundation. These in turn rely on lower level libraries in the UNIX userland, etc. Any Linux desktop works the same way: Gnome/KDE rely on X11 which relies on lower level libraries. You cannot have an X11 library importing a function from Gnome any more than you can have a CoreFoundation library importing a Cocoa function.

Windows does not have this enforced hierarchy. Any lib can import a function from any other, and in practice they do. There is e.g. no way to run a command line only Windows without loading a lot of libraries that were made for the GUI. This means that isolating bugs is very hard, and replacing any part of the system is close to impossible. It was one of the reasons for the Vista disaster. The same thing can be blamed for the problems Apple had with Copland.

MS is (now) well aware that this is a problem, and they have a project called Miniwin that aims to untangle the spaghetti. The image is meant to say that ReFS is untangled. My guess is that NTFS relied on "higher" libraries for things like Unicode text, compression and encryption, and that ReFS solves that by reimplementing such functions at the appropriate level of the stack.

Informative!

[turtle777]

Originally Posted by P 

The wavy vs. flat is a subtle point that I believe they made in the presentation - I have seen it mentioned in passing in reports, anyway. Generally an OS should be organized hierarchically in that there is a kernel and then libraries in layers, with no library relying on a function from a library at a higher level. Apple does this by clearly separating the UI libraries in Cocoa Appkit and Carbon from the underlying libraries in Foundation & CoreFoundation. These in turn rely on lower level libraries in the UNIX userland, etc. Any Linux desktop works the same way: Gnome/KDE rely on X11 which relies on lower level libraries. You cannot have an X11 library importing a function from Gnome any more than you can have a CoreFoundation library importing a Cocoa function.

Windows does not have this enforced hierarchy. Any lib can import a function from any other, and in practice they do. There is e.g. no way to run a command line only Windows without loading a lot of libraries that were made for the GUI. This means that isolating bugs is very hard, and replacing any part of the system is close to impossible. It was one of the reasons for the Vista disaster. The same thing can be blamed for the problems Apple had with Copland.

MS is (now) well aware that this is a problem, and they have a project called Miniwin that aims to untangle the spaghetti. The image is meant to say that ReFS is untangled. My guess is that NTFS relied on "higher" libraries for things like Unicode text, compression and encryption, and that ReFS solves that by reimplementing such functions at the appropriate level of the stack.

Thanks, I learned something new today

-t

[Spheric Harlot]

Ditto!

I'm gonna break the pattern by not quoting the entire post, though.

[turtle777]

-t

[angelmb]

Since I am lost in all this geekness talk, all I see there is that they went from Pepsi to Chevron. That alone, should make Windows a less fatty, well-oiled system.

[Waragainstsleep]

Originally Posted by subego 

Dig your own grave then. The graphic explicitly shows flat is superior to wavy.

[Athens]

As long as what ever replaces HFS+ is as stable as HFS+ I will be happy. I cant count the number of times that NTFS has gone corrupted.

[Waragainstsleep]

I wonder if HFS+ is starting to creak a bit though. Under Lion and Snow Leopard I have increasingly seen the "Mac OS X Cannot repair the volume "XXXXX" copy as much data from it as you can and reformat" message.

[P]

Really? In my experience, NTFS is a solid filesystem. It had journalling long before HFS+ did. Obviously time marches on and in particular ZFS is a much better file system, but I'd put NTFS on a level with the best nineties filesystems (like XFS), and certainly above HFS+ and ext2/3/4. FAT in all its incarnations is terrible in so many ways, but NTFS never let me down.

[besson3c]

Originally Posted by Athens 

As long as what ever replaces HFS+ is as stable as HFS+ I will be happy. I cant count the number of times that NTFS has gone corrupted.

I don't have experience with NTFS, but in relation to every other active filesystem in existence, HFS+ is the epitome of instable. If it were, Disk Warrior would not need to exist.

[besson3c]

Originally Posted by Waragainstsleep 

I wonder if HFS+ is starting to creak a bit though. Under Lion and Snow Leopard I have increasingly seen the "Mac OS X Cannot repair the volume "XXXXX" copy as much data from it as you can and reformat" message.

It's not starting to creak, it's been creaking for years and years now. I can't speak for NTFS, but ext3, ext4, ReiserFS, XFS, ZFS, soon BTRFS, and possibly even UFS are all better file systems and are all available as choices/defaults with other modern OSes. BeFS was probably better too.

HFS+ is hands down the oldest and weakest part of OS X, rivaling the Finder depending on what is most important to you.

[besson3c]

Originally Posted by P 

Really? In my experience, NTFS is a solid filesystem. It had journalling long before HFS+ did. Obviously time marches on and in particular ZFS is a much better file system, but I'd put NTFS on a level with the best nineties filesystems (like XFS), and certainly above HFS+ and ext2/3/4. FAT in all its incarnations is terrible in so many ways, but NTFS never let me down.

Since ext4 has journaled checksums I'd put it above NTFS just in terms of features (I don't know about actual field testing/reliability, performance, etc.), and ext3 may come close to matching it too.

[besson3c]

JFS and Lustre are no doubt better than HFS+ too. Anyway, the point is that it is really showing its age. I will be having sex parties the day it is gone.

[Salty]

What file system does the iPad use?

[jmiddel]

Salty, that is a very interesting question. And having spent 20 mins googling it, I come up with no answer. I googled both iOS 5 and iPad filesystem. Searching for OS X or Windows filesystem serves the answer up in the first page. Am baffled. Any way I found the post below on MacRumors, it addresses a beef I have with iOS:

Originally Posted by CoryTV View Post
..my point is a file system, not print support and drivers.
If we're being technical now... the iPad (and the iPhone and iPod touch) has always had a file system (including having had one yesterday). What you don't have is a file manager or the ability to mount it like a drive on your computer's desktop. They're not the same thing as a file system.
__________________

[besson3c]

Salty, pretty sure it is hfs+ based on the stuff you see at boot time while jailbreaking any iOS device

[Cold Warrior]

Seems specifically hfsx.
ios - What is the file system on iphone? - Stack Overflow
IOS Forensic Analysis: for IPhone, IPad and IPod Touch - Sean Morrissey - Google Books

[Cold Warrior]

Originally Posted by jmiddel 

Salty, that is a very interesting question. And having spent 20 mins googling it, I come up with no answer.

Sadly, google is a shadow of its former self. Results are so poor for very technical searches. Recently they deprecated the + search expression, which was the most important way to search for obscure and technical information.

duckduckgo seems good, it only took me a couple minutes to find references to hfsx. Bing also works with + as expected.

[P]

Originally Posted by besson3c 

Since ext4 has journaled checksums I'd put it above NTFS just in terms of features (I don't know about actual field testing/reliability, performance, etc.), and ext3 may come close to matching it too.

Maybe I was unfair to include ext4 - they have made some real progress there, and I have not been keeping up. ext3 is not in the same league, in my book - it's far slower than ext4 or XFS (no fair comparisons with NTFS, obviously), and the max partition size is 32 TB instead of somewhere in the exabyte range for the others. That alone is enough to make me discount it.

I consider JFS as the same category as XFS and NTFS - a tier 2 filesystem, with ZFS and possibly Btrfs (if it's ever stable) on the top tier. If I'm being generous, HFS+ is comparable to ext3 on tier 3.

[turtle777]

Originally Posted by Cold Warrior 

Sadly, google is a shadow of its former self. Results are so poor for very technical searches. Recently they deprecated the + search expression, which was the most important way to search for obscure and technical information.

Oh they did ? Idiots.
That's why I was having a harder time finding stuff.

-t

[besson3c]

Originally Posted by P 

Maybe I was unfair to include ext4 - they have made some real progress there, and I have not been keeping up. ext3 is not in the same league, in my book - it's far slower than ext4 or XFS (no fair comparisons with NTFS, obviously), and the max partition size is 32 TB instead of somewhere in the exabyte range for the others. That alone is enough to make me discount it.

I consider JFS as the same category as XFS and NTFS - a tier 2 filesystem, with ZFS and possibly Btrfs (if it's ever stable) on the top tier. If I'm being generous, HFS+ is comparable to ext3 on tier 3.

Well, ext3 supports LVM/live partition resizing, which right there puts it over HFS+ in my book. I believe ext3 also has a block level journal, whereas HFS+ has a file based journal?

[Cold Warrior]

Originally Posted by turtle777 

Oh they did ? Idiots.
That's why I was having a harder time finding stuff.

-t

It's been very frustrating. There's an epic thread somewhere on the google forums about this change. A google employee chimed in early on and talked about how "" quotes are the same as +, then other posters exposed that as incorrect. Google has apparently decided it knows best. Quotes and plus are very different, and I could use them in combination with - (minus) to go into the internet with a scalpel and find relevant results. Now I can't — at least with google. I still google for basic things, but anything with any complexity, I use duckduckgo.

[mduell]

Originally Posted by besson3c 

- Using iCloud for storing checksum data copies of files for self heal so that single disk systems never have to deal with data corruption (no more fsck, Disk Warrior, etc.) This might take some time before it is cost effective for Apple and so that this works well without requiring a whole lot of bandwidth, but I think we'll eventually get there.

You can't heal from a checksum, you can only figure out that you're corrupt. Also this crosses a lot of lines in the layer model P discussed above.

Originally Posted by besson3c 

- deduplication

Typical users don't need this/it wont be useful, and don't want the performance side effects that come with it on consumer-grade equipment.

[besson3c]

Originally Posted by mduell 

You can't heal from a checksum, you can only figure out that you're corrupt. Also this crosses a lot of lines in the layer model P discussed above.

I know, that's why I said "copies of files", but I should have said that iCloud would store "parity data", not "checksum data".

Typical users don't need this/it wont be useful, and don't want the performance side effects that come with it on consumer-grade equipment.

Fair enough

[P]

Originally Posted by besson3c 

Well, ext3 supports LVM/live partition resizing, which right there puts it over HFS+ in my book. I believe ext3 also has a block level journal, whereas HFS+ has a file based journal?

Well yes, but it doesn't have extents or delayed allocation, two features that were added in ext4 and that already exist in HFS+. It also has more metadata and logging features. Swings & roundabouts.

(Not sure about the journaling system for HFS+. I know that efs3 can be configured to journal ALL writes, not only metadata writes,but it is rarely used as the performance hit is...substantial, shall we say)

[mduell]

Originally Posted by besson3c 

I know, that's why I said "copies of files", but I should have said that iCloud would store "parity data", not "checksum data".

Why not just store a copy of the important files in the cloud instead of trying to wedge some insane cloud-synced parity scheme into the filesystem?

[Big Mac]

What are important files?

[besson3c]

Originally Posted by mduell 

Why not just store a copy of the important files in the cloud instead of trying to wedge some insane cloud-synced parity scheme into the filesystem?

I think there are a number of ways in which iCloud could be potentially utilized, but my original thinking was that it could be used to take the risk out of relying on a single drive, so long as you have internet access. A backup would help, but realtime healing would be even cooler.

[mduell]

Originally Posted by Big Mac 

What are important files?

TBD, by Apple or the user, but I accept that some files are more important than others and possibly worth uploading to the cloud.

Originally Posted by besson3c 

I think there are a number of ways in which iCloud could be potentially utilized, but my original thinking was that it could be used to take the risk out of relying on a single drive, so long as you have internet access. A backup would help, but realtime healing would be even cooler.

Parity data is worthless if you only copy of the original data is on a dead single drive.

I suppose you could break a block into 5 chunks, calculate the parity, upload the parity to the cloud. Then if one (and only one and you know which one) chunk becomes corrupt, you download the parity and recreate the chunk. A horrifically complicated mess to achieve the same result as copies=2 with slightly better storage efficiency.

A vary narrow case where parity in the cloud is useful to a single drive system. Oh and you need network access before another chunk goes corrupt.

[besson3c]

Originally Posted by mduell 

Parity data is worthless if you only copy of the original data is on a dead single drive.

I suppose you could break a block into 5 chunks, calculate the parity, upload the parity to the cloud. Then if one (and only one and you know which one) chunk becomes corrupt, you download the parity and recreate the chunk. A horrifically complicated mess to achieve the same result as copies=2 with slightly better storage efficiency.

A vary narrow case where parity in the cloud is useful to a single drive system. Oh and you need network access before another chunk goes corrupt.

Well, it would deal with file corruption, not a failing disk. I'd envision a number of internal write failures triggering an OS notification that the disk is dying (which in and of itself would be a great feature), and something like this trying to prevent permanent data loss as best as it can during this time period. I don't know how often files become corrupt on a perfectly healthy disk.

Perhaps the totality of this is impractical, but it will still be nice for Apple to have these tools available available.

Perhaps the best replacement for a single SATA drive or SSD would be a bunch of very small size/capacity ones in a RAID 5/10 configuration?

[mduell]

It's an elaborate solution to a narrow problem.

You can get your OS notification of read/write failures without any cloud/parity tie in.

More disks in each client has obvious benefits, but contributes substantially to cost, size, failure rate, etc.

[rickey939]

Appears very user-friendly.

[besson3c]

Originally Posted by mduell 

It's an elaborate solution to a narrow problem.

You can get your OS notification of read/write failures without any cloud/parity tie in.

More disks in each client has obvious benefits, but contributes substantially to cost, size, failure rate, etc.

Well, as far as costs go, is it fathomable that there would be, say, 3 Flash smaller sized drives that equals the cost of one? Is it fathomable that they fit these three flash drives in the space of one 2.5" SATA drive, that the failure rate of 3 flash drives would be less than one SATA drive, and that all of this would be cheaper than one fairly large SSD?