Benchmarking The Experimental Bcachefs File-System Against Btrfs, EXT4, F2FS, XFS & ZFS

[tildearrow]

The logarithmic result overview (last chart on last page) is amazing!

For some reason it reminds me of Gyruss.

[chilinux]

Oracle has nothing to do with Open Source ZFS. ZoL or BSD's ZFS.

Unless ZoL does a clean room style ground up re-implementation, it seems like only Oracle can provide a true solution to getting ZFS under the GPL.

[tajjada]

Unless ZoL does a clean room style ground up re-implementation, it seems like only Oracle can provide a true solution to getting ZFS under the GPL.

No ... even Oracle cannot.

The only way to legally change the license is to get *all* copyright holders to agree on the change. You cannot change the license for something you do not hold the copyright to. The copyright is held by the different contributors (companies or individuals), each for their own code. Even if Oracle (or any other party involved in the development of the project) wants to change ZFS to the GPL, all other contributors must agree too. It is practically impossible to do.

EDIT: of course, the license can also be changed illegally without contacting all the copyright holders, but then you risk getting sued by those who haven't agreed. Try to contact as many people as possible, hoping that everyone who actually cares will respond, and if all is good, change the license and hope that those who haven't responded don't care and won't sue (or make sure you have a way to delete/undo their contributions). IANAL but i think this is how it goes.

[k1e0x]

Unless ZoL does a clean room style ground up re-implementation, it seems like only Oracle can provide a true solution to getting ZFS under the GPL.

Who cares about the GPL? CDDL/MPL is better anyway. More permissive. Let the distro's packages it like an Nvidia driver. Done, apt-get install zfs.. problem solved.

tajjada You are exactly right sir. Good info to share. ZoL contributions are owned by the ZoL developers. (and it's more than 50% new code not found in Oracle ZFS as far as I hear, Oracle wishes they had ZoL features in their version.. they cant get them.)

[chilinux]

Who cares about the GPL? CDDL/MPL is better anyway. More permissive. Let the distro's packages it like an Nvidia driver. Done, apt-get install zfs.. problem solved.

tajjada You are exactly right sir. Good info to share. ZoL contributions are owned by the ZoL developers. (and it's more than 50% new code not found in Oracle ZFS as far as I hear, Oracle wishes they had ZoL features in their version.. they cant get them.)

I can find plenty of articles indicating lack of a GPL version of ZFS is impacting adoption. Claiming a file system being installed via package management is just the same situation as installing binary blob video drivers really misses a key point to advocating for ZFS. All the programs already installed will take advantage of the nvidia video driver just simply by rebooting. The same can't be said of a root file system still being on XFS or EXT4 after a reboot. These ZFS benchmarks leveraged Ubuntu's feature to support installing a ZFS root file system. But Canonical has a history of not being fully invested in the long term with their key design decisions. Just because Canonical indicated they fully stand behind Unity and Mir in the future does not mean much today. Likewise, just because Canonical indicate they fully stand behind ZFS today may not mean much tomorrow. Advocating for ZFS should at some point include addressing the biggest issue impacting more widespread adoption. There are plenty of articles indicating that the license is that biggest issue right now if even you want to just write it off as "trolling."

As to Oracle not having the legal right to change the license, I would agree that they don't control all the code in ZoL--I wasn't talking about ZoL for now. For all of ZFS for Solaris, every contributor signed a joint ownership contributors agreement giving Sun/Oracle the rights to claim shared ownership of the code and being able to relicense. In terms of Oracle not being able to release under the GPL for the original ZFS code because of the work involved in getting everyone to agree, that simply is not true. They already have the agreements and the legal rights to release all of ZFS for Solaris code under the GPL if they wanted, they just have decided not to.

In terms of ZoL's contributions, they have made several important changes but the largest contributor to the entire code base still seems to be Sun/Oracle. A review of the changelog and git commits seem to show a reasonably small number of major contributors. If Oracle agreed to the foundational code base to be under dual GPL/CDDL, it shouldn't be any more unrealistic for ZoL to also change to a dual license than it was for the Wine program to go from BSD to LGPL. But it still seems like that process would have to start with Oracle.

[k1e0x]

I can find plenty of articles indicating lack of a GPL version of ZFS is impacting adoption. Claiming a file system being installed via package management is just the same situation as installing binary blob video drivers really misses a key point to advocating for ZFS. All the programs already installed will take advantage of the nvidia video driver just simply by rebooting. The same can't be said of a root file system still being on XFS or EXT4 after a reboot. These ZFS benchmarks leveraged Ubuntu's feature to support installing a ZFS root file system. But Canonical has a history of not being fully invested in the long term with their key design decisions. Just because Canonical indicated they fully stand behind Unity and Mir in the future does not mean much today. Likewise, just because Canonical indicate they fully stand behind ZFS today may not mean much tomorrow. Advocating for ZFS should at some point include addressing the biggest issue impacting more widespread adoption. There are plenty of articles indicating that the license is that biggest issue right now if even you want to just write it off as "trolling."

As to Oracle not having the legal right to change the license, I would agree that they don't control all the code in ZoL--I wasn't talking about ZoL for now. For all of ZFS for Solaris, every contributor signed a joint ownership contributors agreement giving Sun/Oracle the rights to claim shared ownership of the code and being able to relicense. In terms of Oracle not being able to release under the GPL for the original ZFS code because of the work involved in getting everyone to agree, that simply is not true. They already have the agreements and the legal rights to release all of ZFS for Solaris code under the GPL if they wanted, they just have decided not to.

In terms of ZoL's contributions, they have made several important changes but the largest contributor to the entire code base still seems to be Sun/Oracle. A review of the changelog and git commits seem to show a reasonably small number of major contributors. If Oracle agreed to the foundational code base to be under dual GPL/CDDL, it shouldn't be any more unrealistic for ZoL to also change to a dual license than it was for the Wine program to go from BSD to LGPL. But it still seems like that process would have to start with Oracle.

Not sure if Oracle even can contribute to the code without breaking the CDDL (if they even would, lol ya right) They fired Mark Maybee when he spoke at a OpenZFS conference. https://twitter.com/bcantrill/status...351360?lang=en

But perhaps "Oracle will step in and save ZFS.. and make it GPL and lovey dovey." Enjoy your pipe dream.

Also ZFS has more adoption than any of these filesystems sans EXT4. (and maaybe mayybe XFS)

[k1e0x]

And you do realize that a formatted ZFS storage array on the metal is often used to host RedHat virtual machines formatted to XFS. The "adoption" is in layers but the big data is on ZFS.

ZFS doesn't compete with XFS, it competes with NetApp.. Sure you can run ZFS on your desktop and your backups.. but when you search for "XFS based storage appliance" Google thinks it's a typo and says "Did you mean ZFS based storage appliance" ;p

[Dedobot]

Sync to disabled and ZFS beats them all
Seriously I'm surprised how well both zfs and btrs stand to others . The copy on write nature insist a lot of background actions .
Give them RAM , good number of vdevs , the best ZIL money can by and COW will shine.

[Quetzalcoatlus]

I had a great experience with it in this very strange kind of MacGuyvered system i have in place for rendering openstreetmaps out of the whole planet database, but without nearly enough SSD storage to fit the whole thing at all unless heavily compressed. Clocks in about 900GB and I can only spare 350.

So clearly that eliminates any filesystem without compression. Lz4 on zfs gets it down to a little over either 300 not much head room; zstd gets it down about another 10% to 270. I haven't yet had a chance to see if there's much difference in DB performance, but bcache blew the doors off the other two when it came to copying the database off of spinning disk hard storage: was pushing about 100 MB/s over USB 3 no Iess. (Of zstd compressed pgsql tables)

Copy to btrfs much much slower 20-40% no matter how low i set the compression so it may have more weird corner cases where it just turns out absurdly fast like that. I wasn't convinced it was copying the data at all until it started spitting out maps afterwards!

[Quetzalcoatlus]

Not sure if i can edit actually it may have been the magnetic spinning disk itself maxing out and slowing down bcache; btrfs was clearly the limiting end on that copy