No announcement yet.

Benchmarks Of The Official KQ ZFS Linux Module

  • Filter
  • Time
  • Show
Clear All
new posts

  • #16
    Originally posted by deanjo View Post
    I'm not sure how you see it as a violation of the GPL. The source is available. The GPL does not prevent the supplying of binaries as well.
    It can. It depends on the binaries.

    If this was the case then pretty much everything that utilized linux (including binary delivered distributions) would be in violation themselves.
    The #1 most important thing to remember about the GPL is that it is a copyright license. The scope of the copyright license is restricted by copyright law.

    There is a very important term in copyright law called 'derivative work'. It's a legal term that is designed to define what happens when you create a new work by taking a old work and adding to it. It's defined by the law and case precedent, not licenses like the GPL.

    The GPL basically says that if you distribute derivative work under the GPL license then you are required to make the source code available under the terms of the GPL.

    Take, for example, the Nvidia kernel drivers. The Nvidia drivers were not original written for Linux. They are more then likely Windows driver code with some compatibility added to make it work with Linux. So therefore they are probably not derivative work and thus are not covered by the GPL. No matter how much the Linux developers or GNU or whoever does not like it they are not the ultimate authority here.

    But if you compile the drivers then that act causes portions of Linux code to be sucked into the actual resulting binaries. If that is true then binaries can violate the license.

    The #2 most important thing to remember is that law is NOT software. That is it's not like code that you write and then that is what sets up the rules. You can read all the case law and understand the license and read USA copyright code... but that is not enough. You cannot take the law literally. It must be interpreted.

    That is one Judge may say that Nvidia binaries are derivative and violate the GPL... and another Judge may say that it does not. And BOTH can be right and their decisions can both be 100% legal and correct.

    This is how the law works.

    The #3 important thing with the GPL is that it places no restrictions on usage. ONLY distribution. You can combine GPL'd code with all the closed source code you want and use it and it all is 100% legal.. because the license allows this. You have no requirements for sharing code or anything like that. The restrictions only kick in when you want to distribute the software.

    So in this specific case the ZFS driver code is a derivative of the Solaris kernel code and is licensed under the CDDL. Just because you make it compatible with Linux does not make it derivative of Linux. You actually have to use Linux code to make it derivative. If they don't do that then it's 100% legal to distribute.

    Once you compile it then it'll suck in portions of the Linux kernel code and that is when you _CAN_ run into problems.

    It is absolutely true that some binary drivers can violate copyright, but others may not. It depends on the details.


    • #17
      Wow, I am really surprised be good performance. Would need to check performance on my 32bit multidisk Debian box. But it looks would be much better than current zfs-fuse (which isn't so bad, but had some problems).

      Anybody knows if NFS exports will work on this zfs?


      • #18
        Originally posted by Hugh View Post
        How can KQ Infotech be authorized to distribute binaries of the code? The binaries are surely derived in part from GPLed code (the kernel) and thus have the GPL apply to the whole work (including the ZFS code) OR distribution is not authorized.

        This is precisely why distros cannot incorporate the ZFS code.

        As far as I understand it, what might well be authorized is the distribution of the ZFS port's source code so that individual users could build a binary for their own system. But those users cannot distribute the resulting binary either.

        Perhaps the GPL doesn't work in India. But I doubt it: their law is most likely like our law (Canada), both being descended from British law and the Berne Convention.
        There is a provision to the linux licence specifying what kernel modules are allowed not to be redistributed under the GPL. Some linux symbols are exported as GPL-only. To use them in a module, the module has to be GPL. Some symbols are allowed to be used by non-gpl modules, similar to the way kernel API is used.
        Maybe KQ Infotech are not using the GPL symbols and thus doing what the binary blob developers do.
        Is it possible to implement a file system without using the GPL symbols?


        • #19
          4k sectors WD EARS and LLNL zfs on linux

          I would really like to see this benchmark on some 4k sector drives such as the WD EARS. I tried it on my system and the zpool automatically detects 512 byte sectors which are emulated by the drives which results in terrible performance.

          Is there a way to force 4k sectors in linux, similar to the BSD gnop device? This would only be needed at the zpool creation time.


          • #20
            Originally posted by Proximo View Post
            Please, test the ZFS Linux port in different software RAID versions (RAID1,RAID1+0 and especially RAID-Z vs RAID5. ZFS should shine in multidisk configurations.
            Take a look at this article to see what is really capable ZFS -->
            I second this. Actually, I really just want to know if Raid-Z is supported in this release.


            • #21
              Arc2 & zil

              I've been fascinated by ZFS ever since reading all the benchmarks at Would it be possible for you to run some benchmarks with mechanical storage and SSD ARC2 and SSD ZIL drives for speed?

              Preferably on moderately decent hardware:
              64-bit CPU
              8 GB ram
              6 Gb sata ports

              As others have mentioned, testing RAIDZ and mirroring would also be awesome.


              • #22
                Originally posted by kneufeld View Post
                I've been fascinated by ZFS ever since reading all the benchmarks at Would it be possible for you to run some benchmarks with mechanical storage and SSD ARC2 and SSD ZIL drives for speed?

                Preferably on moderately decent hardware:
                64-bit CPU
                8 GB ram
                6 Gb sata ports

                As others have mentioned, testing RAIDZ and mirroring would also be awesome.
                Look below, I provide a link where they bench BTRFS vs ZFS, and ZFS outperforms BTRFS, because BTRFS can not handle many discs. In other words, BTRFS, scales bad.

                Originally posted by RealNC View Post
                I've just read the LKML thread. I don't see how it's broken. The devs replied to Edward explaining why his concerns don't affect BTRFS. (And there are no further replies from Edward after that.)
                Ok, you have not read the BTRFS mail lists nor forums. BTRFS is quite unstable and have been massively critizised, not only by Edward.

                Originally posted by drag View Post
                That is because BTRFS is programmed in C code, while ZFS is written in magic pixie dust.
                Well, I told you that ZFS scales better than BTRFS, and it is because of ZFS is targeted to large scale Enterprise, whereas BTRFS, ext4, ... are targeted for small servers. It is not bullshit, nor FUD. I can prove it to you (as always). Read this and see how bad BTRFS performs:

                Originally posted by BlackStar View Post
                He's probably set up google alerts for 'btrfs' and 'zfs' and posts spam each and every appearance of these words.

                It doesn't help that his posts are pure FUD at best.
                You clearly dont know what FUD is. I suggest you read about FUD, beacuse you claim that I FUD.

                I provided benchmarks which shows that BTRFS scales bad, compared to ZFS on many discs. I did not lie about that. This proves I do not FUD, I speak true, and this proves that Blackstar speaks not true about me, I am not a FUDer (because I showed proof). Or, can you Blackstar provide links as I did, that supports your claim about me? No? Then dont call me a FUDer, because that would make you the FUDer, when you say untrue things about other people.


                • #23
                  Please include more oldies

                  As usual, I wish there were more of the "old" filesystems included, such as JFS and Reiser3, but I guess it's nice you included XFS instead of the article being 100% dedicated to the next-generation stuff.

                  I know the oldies don't really compare to the new stuff in terms of features (where ZFS particularly shines), but there are still a lot of cases where I think they can outperform, especially given that you're benchmarking on a single-disc setup. But thinking ain't knowing, which is why it'd be cool to see the numbers.

                  I wanna know if the new guys are ready to take on reiser3's awesomeness for maildirs, JFS' quickness at doing certain things with big files particularly on cpu-challenged systems like Atom, and so on. By leaving out some of the key players, a person still can't look at a benchmark article like this and guess which filesystem is best for a particular job.

                  Aside from that complaint, though, this is a pretty interesting article. I didn't know if ZFS (once freed of FUSE) would really be a competitive performer and it looks like there are some cases where it really is. Exciting.


                  • #24
                    RaidZ in use here

                    FYI, I have 4 2TB drives set up on my gentoo htpc box in a raidz configuration. It's working pretty well. I've done some basic benchmarking (link is ). The write speed isn't great but so far the filesystem seems stable which was my entire reason for wanting to go with zfs for softraid.


                    • #25
                      Very interested in raidZ multidisk benchmarks. Bonus points for using SSD for zil and arc2 cache as others have mentioned.