Benchmarking Debian's GNU/kFreeBSD

[trasz]

ZFS deliberately chose to sacrifice some freedom by using the CDDL license (http://en.wikipedia.org/wiki/CDDL):

Still, the CDDL license is Free Software (according to FSF) and Open Source (according to OSI) and is less restrictive than GPLv2, let alone GPLv3 - for example, it doesn't prohibit one from linking with code under any other license, which is what GPL does.

This is also why Debian probably never will include ZFS "natively" anyhow, but rather upgrade their default file system to ext4, which btw was good enough for Google, as noted by the trolling Stephen J. Vaughan-Nichols, in "The best Linux file system of all?" http://blogs.computerworld.com/15413..._system_of_all and who quoted Google's Michael Rubin, a senior staff engineer:



So, despite all the hype of ZFS, maybe Google had their reasons.

Google has very different requirements compared to the usual server uses - for example, their data is replicated between machines, so reliability of a single machine is less important - and reliability (snapshots, checksums, built-in RAID etc) is a very important part of ZFS. Also, I don't think Google needs all the flexibility ZFS provides. They just want to replace 1980's technology (ext2 is basically a simplified version of the original UFS, released in 1984, without most of the features UFS got during the following 20 years) with something more 1990's (extents, journalling, defragmentation - VxFS had all this over a decade ago).

[BuissonVert]

Hello,

In the article, you wrote a couple of time that Debian Squeeze would be released in March. This is most probably wrong, as March 2010 is only the beginning of the Testing Freeze. The release usually comes a few months after the freeze.

http://lists.debian.org/debian-devel.../msg00002.html

[Apopas]

Google's choice of ext4 is rather interesting, particularly given the serious performance regressions in the more recent kernels.

This is just a matter of the default configurations of Ubuntu.

[Darkfire Fox]

Is this the reason why Debian-stable is a year behind everyone else?

**shot**

Not only do they support two dozen arcitectures, but now they'll have three kernels?

(for the ill-humored, I'm using hyperbole)

[droidhacker]

- Last time I checked Hurd was supposed to be the next generation über-kernel, so investing money/time in FreeBSD must have a good reason.

Hurd... in development since 1990.
20 years and there is YET to be a stable release.
BSD has been around and stable since.... about the beginning of time.

If they want a kernel option besides Linux, BSD is ready and waiting. Debian's been going around in circles with Hurd since they tried to pick it up back in 1998.


Seriously, I'll believe in Hurd when it actually becomes more than vaporware.

[kraftman]

Still, the CDDL license is Free Software (according to FSF) and Open Source (according to OSI) and is less restrictive than GPLv2, let alone GPLv3 - for example, it doesn't prohibit one from linking with code under any other license, which is what GPL does.

Glad to see you're back. GPL is more restrictive and thus it's probably the best for projects which are direct competitors to projects that use proprietary or bsd licenses. GPL is to protect the code not to allow people to do what they want with it.

[MaestroMaus]

Hurd... in development since 1990.
20 years and there is YET to be a stable release.
BSD has been around and stable since.... about the beginning of time.

If they want a kernel option besides Linux, BSD is ready and waiting. Debian's been going around in circles with Hurd since they tried to pick it up back in 1998.


Seriously, I'll believe in Hurd when it actually becomes more than vaporware.

It's probably there. I think someone can post you some code-screenshots (it's opensource afterall). The only problem is work on it is too damn slow.

[smitty3268]

Google's choice of ext4 is rather interesting, particularly given the serious performance regressions in the more recent kernels. I wonder if they're not just lining themselves up for a world of hurt.

I don't know how many times this can be said, those ext4 performance regressions DON'T EXIST if you don't enable all the new safety features being added. I'm quite certain google will fine-tune the FS for their needs, and I wouldn't be surprised to see them completely disable the journal, for example, which is something ext4 allows.

They've done benchmarking for their choice, but very importantly point out that it's the best file system for their particular needs and intended use. Doesn't make it the best file system every need.

I think they pretty much came out and said the reason they chose ext4 over some of it's competition was because they could do an online upgrade, so switching to a completely new OS as well as a FS probably wasn't considered too closely if they were really that concerned about uptime.

[StringCheesian]

I don't know how many times this can be said, those ext4 performance regressions DON'T EXIST if you don't enable all the new safety features being added. I'm quite certain google will fine-tune the FS for their needs, and I wouldn't be surprised to see them completely disable the journal, for example, which is something ext4 allows.

Disabling barriers isn't enough:

When using the nobarrier option the number of transactions per second had nearly doubled. However, even when using this EXT4 mount option, the 222 TPS that it had produced is still significantly less than 560~598 TPS that can be found in pre-2.6.32 kernels. In other words, there are still significant performance penalties in the more recent kernels.

[david.given]

Some of those filesystem benchmarks look very dubious to me --- UFS shouldn't be that much slower than ext3. Can you confirm that you did have SoftUpdates turned on? Without that, you'll cripple the FreeBSD file system performance.