DragonFlyBSD Improves Performance Against Linux

[gamerk2]

I someone who regularly benchmarks clang/llvm vs GCC I can say I've not seen any results (atleast not as recent as this year, before that my memory may fail me) where clang/llvm beats GCC where I use -O3, which is the optimization level where the compilers strive to produce the fastest code. And as we can see from the results you linked to, whenever -O3 is used GCC is 'superior' to clang/llvm, and also in most of the cases where no optimization level is set (which really doesn't count though as it is worthless) and the only times clang/llvm beats GCC is when there is no -O optimization level which means GCC defaults to -O0 which is 'no optimization' and aimed for debugging.

Typically my tests results in 5-20% better performance with GCC versus clang/llvm, and when I've tried the GCC 4.8 versus LLVM 3.2 snapshots GCC has actually increased the performance gap (that said, snapshots are anything but conclusive). It should be noted that my benchmarks are all on x86_64, I have no idea of how x86 or for example ARM architectures compare.

On another note I find Dragonfly most interesting from a technical standpoint, keep up the good work!

Think of it this way: GCC is faster now, but it will be far harder to make any significant long term performance enhancements, because its design is more or less set in stone. LLVM/CLANG will be significantly faster in a few years, and have better toolkit support to boot.

[froznen]

Only two posts and a big fail. You linked to unreleased version. It will be nice to see some newer Linux distro rather than old SL. Btw. I always wondered why it'sFreeBSD pushed so hard rather than DragonFlyBSD. The later is not only faster, has great community, but it's also much more innovative. In contrary freebsd community are just envy fanboys and trolls.

It would be interesting to see newer Linux Distro's benchmarked, but SL is a good comparison. In terms of server operating systems, the SL tested is basically present day. Debian and anything compiled with Redhat(CentOS and SL) all use similar kernels. You don't see too many (if any) popular Linux server Distros use the newer kernels.

[smitty3268]

I someone who regularly benchmarks clang/llvm vs GCC I can say I've not seen any results (atleast not as recent as this year, before that my memory may fail me) where clang/llvm beats GCC where I use -O3

Well, if you compare GCC 4.4 to LLVM 3.2 it might be different.

[Sergio]

Only two posts and a big fail. You linked to unreleased version. It will be nice to see some newer Linux distro rather than old SL. Btw. I always wondered why it'sFreeBSD pushed so hard rather than DragonFlyBSD. The later is not only faster, has great community, but it's also much more innovative. In contrary freebsd community are just envy fanboys and trolls.

Why are you always bitching about FreeBSD/Solaris/insert-your-non-Linux-OS-here? (I'm surprised you have something kind to say about DragonFlyBSD).
You criticize FreeBSD for having a fanboy comunity, but you always sound like a fanboy yourself.

[kraftman]

Yes you're right. Linux rox, and has no scalability issue, it's well known. And postgresql scales easily on latest linux kernels. Oh wait ! http://http://lwn.net/Articles/518329/

Only two posts and a big fail. You linked to unreleased version. It will be nice to see some newer Linux distro rather than old SL. Btw. I always wondered why it'sFreeBSD pushed so hard rather than DragonFlyBSD. The later is not only faster, has great community, but it's also much more innovative. In contrary freebsd community are just envy fanboys and trolls.

[XorEaxEax]

Also - performance tests indicate that clang (which FreeBSD will be switching to in 10.x) - is superior to gcc
in some performance tests, but slower in others - so in other words, they are extermely competitive:

http://www.phoronix.com/scan.php?page=news_item&px=MTE5OTM

I someone who regularly benchmarks clang/llvm vs GCC I can say I've not seen any results (atleast not as recent as this year, before that my memory may fail me) where clang/llvm beats GCC where I use -O3, which is the optimization level where the compilers strive to produce the fastest code. And as we can see from the results you linked to, whenever -O3 is used GCC is 'superior' to clang/llvm, and also in most of the cases where no optimization level is set (which really doesn't count though as it is worthless) and the only times clang/llvm beats GCC is when there is no -O optimization level which means GCC defaults to -O0 which is 'no optimization' and aimed for debugging.

Typically my tests results in 5-20% better performance with GCC versus clang/llvm, and when I've tried the GCC 4.8 versus LLVM 3.2 snapshots GCC has actually increased the performance gap (that said, snapshots are anything but conclusive). It should be noted that my benchmarks are all on x86_64, I have no idea of how x86 or for example ARM architectures compare.

On another note I find Dragonfly most interesting from a technical standpoint, keep up the good work!

[staalmannen]

[QUOTE=dfcat;290858]"

I am curious to know... would a DragonFly vkernel be able to run in user land of another OS? It is considered similar to UML but I wonder if it also could be used similarly to coLinux.
"

The vkernel requires some host-side support - specifically ability to create 'vmspaces' which are separate virtual memory maps,
and also a some minor hooks to interface to hardware / host side resources. Doable, but alot of work. Someone interested in knowing
this level of detail would be better of giving DF hardware assisted virtualization so you could run DF and virtualize the 'other' os where
needed

UML iirc also required same.

Probably the best bet right now to run dragonfly on another OS is virtualization e.g. via KVM / QEMU / VirtualBox.
I am personally, other than on several native dragonfly machines, running df on VirtualBox/Linux, with some devs using KVM.

Thanks for the clarifications

Personally I just find chroots/jails in directories so much more convenient than typical virtualization and disk images and I think it would be really cool to be able to boot a BSD system directly on top of a Linux kernel (for all that pesky hardware that may not want to work otherwise)

Just another question - how does the vkernel compare to NetBSD RUMP? It can be run on different OSes (although I have failed building it on my Linux system) but on the other hand, it does not support running applications but rather provides various kernel features like file systems etc.

[Sergio]

So DragonFlyBSD is very interesting and definitely not a waste of time...

[soupbowl]

Nice to see dragonfly bsd doing well, I have been watching it for years.

[dfcat]

"
[QUOTE=staalmannen;290857]I am curious to know... would a DragonFly vkernel be able to run in user land of another OS? It is considered similar to UML but I wonder if it also could be used similarly to coLinux.
"

The vkernel requires some host-side support - specifically ability to create 'vmspaces' which are separate virtual memory maps,
and also a some minor hooks to interface to hardware / host side resources. Doable, but alot of work. Someone interested in knowing
this level of detail would be better of giving DF hardware assisted virtualization so you could run DF and virtualize the 'other' os where
needed

UML iirc also required same.

Probably the best bet right now to run dragonfly on another OS is virtualization e.g. via KVM / QEMU / VirtualBox.
I am personally, other than on several native dragonfly machines, running df on VirtualBox/Linux, with some devs using KVM.