Running FreeBSD 12.0 With Intel Xeon Scalable Cascade Lake / Gigabyte S451-3R0 Server, Benchmarks Against Linux

[phoronix]

Phoronix: Running FreeBSD 12.0 With Intel Xeon Scalable Cascade Lake / Gigabyte S451-3R0 Server, Benchmarks Against Linux

Over the past week we've been delivering several benchmarks of the new Intel Xeon Scalable "Cascade Lake" processors under Linux with the Platinum 8280 processors. The Linux support and performance is hitting expectations, but what about on the BSDs? I spent some time this week trying out the dual Intel Xeon Platinum 8280 processors within the Gigabyte S451-3R0 Storage Server. The FreeBSD 12.0 support has been in great shape and in this article are some comparison benchmarks between various Linux distributions and FreeBSD for those curious.

http://www.phoronix.com/vr.php?view=27742

[edwaleni]

Netflix will like those transcoding numbers. They are a big FreeBSD shop for their CDS.

[tildearrow]

I have a question. Why do you use BLAKE2 for comparing different OS'es but why do you never use it for comparing different CPU's? Because the results are in *Cycles* per byte, I'd say it makes more sense to compare between different CPU's than different OS'es.

[Michael]

I have a question. Why do you use BLAKE2 for comparing different OS'es but why do you never use it for comparing different CPU's? Because the results are in *Cycles* per byte, I'd say it makes more sense to compare between different CPU's than different OS'es.

Mostly from having so many tests and occasionally forgetting about them... In this case remembering about blake2 and that I hadn't tried it out on cascadelake yet, so tossed it in.

[Snaipersky]

What I'm getting from this is OSL is a damn fine server OS.

[starshipeleven]

What I'm getting from this is OSL is a damn fine server OS.

Which means that SUSE Enterprise Linux should also be, as Leap is basically the free version of it.

[Wojcian]

Those results are devastating for FreeBSD. It seems Clear Linux is the best choice for transcoding and encoding files. Clear winner.

[k1e0x]

The filesystem benchmarks suffer because ZFS needs it's ARC to be performant in them. It's tuned for real world use cases not synthetic tests. It depends what exactly you're testing, ZFS does a lot more work that most other filesystems.. so if you're testing kernel IO you may want UFS. (Netflix uses UFS). Transcoding states look pretty good to be honest on a default FreeBSD compared to a highly optimized Clear Linux. (Netflix does ton's of changes on their platform as well)

I can assure you that ZFS on FreeBSD will smoke ZFS on Linux with the new Linux 5.0 changes.

[Wojcian]

The filesystem benchmarks suffer because ZFS needs it's ARC to be performant in them. It's tuned for real world use cases not synthetic tests. It depends what exactly you're testing, ZFS does a lot more work that most other filesystems.. so if you're testing kernel IO you may want UFS. (Netflix uses UFS). Transcoding states look pretty good to be honest on a default FreeBSD compared to a highly optimized Clear Linux. (Netflix does ton's of changes on their platform as well)

UFS is old and boring horse, so thanks, but no. There are many real world use cases where ZFS sucks. Yes, Clear Linux is highly optimized, but Ubuntu has debugging turned on by default.

I can assure you that ZFS on FreeBSD will smoke ZFS on Linux with the new Linux 5.0 changes.

That's funny, because ZFS on Linux is running in userspace, but it's still in much better shape than FreeBSD implementation. I mean it has lower performance, but from some reason it's being treated like the first class citizen. I think we can safely assume FreeBSD have no resources to invent or even maintain more interesting projects. At first it took ZFS from Solaris and now it takes from ZoL.

[k1e0x]

That's funny, because ZFS on Linux is running in userspace, but it's still in much better shape than FreeBSD implementation. I mean it has lower performance, but from some reason it's being treated like the first class citizen. I think we can safely assume FreeBSD have no resources to invent or even maintain more interesting projects. At first it took ZFS from Solaris and now it takes from ZoL.

It's not, you're thinking of ZFS-FUSE, an old (dead?) implementation.

I've always found it funny that Linux subscribes to the fallacy that more people automatically produce better designs but in the real world you often see the opposite. An you are right BSD does not have a NIH complex.. (and if you want to get technical about who made things first.. Most OS's owe quite a lot to the BSD's)

They simply just want to make the best system they can, they don't have fragile ego's about who's idea is was first.