The Performance Of Ubuntu Software Running On Windows 10 With The New Linux Subsystem

[chithanh]

To be a fair comparison we would need to have an additional windows FS benchmark (without any Linux layer)

Not necessarily, comparing ntfs-3g with another FUSE&non-FUSE filesystem would be enough to find out how much overhead there is on Linux. Then we'd know which slowness is due to Windows instead of due to NTFS.

[Volodik]

Here found a way to run graphical applications and even entire DE, but not performance tested. Now you can test them.
P.S. Sorry, English is not the native language for me

[rasmithuk]

I'm not talking desktop OS, I'm talking Server. Microsoft does not have a proper LVM even on their Server OS's. What LVM are you talking about?? Because if it's Windows Disk Manager.... LOL... just LOL. That's like a speak-n-spell My First LVM. It's so bad, Veritas offers their VxFS and Veritas LVM for Windows servers and charges $thousands for it. Windows Disk Manager does not do ANY of the core functions of a real LVM. Can you grow your C:/ live on the fly no downtime? No. Can you migrate an LV from one PV to another? No. Can you do all kinds of creative striping, mirroring, and spanning? Nope, you can only do a 2-member mirror OR a 2-member stripe (but not both) and creating either of them is a one-way function. So please, what is this Windows LVM you speak of, because this bad joke called Windows Disk Manager can't possibly be it.

Assuming you're using Storage Spaces (added in Server 2012 and 8)...

Can you grow your C:/ live on the fly no downtime?

Sort of. The boot drive can't be on a storage space, but any other volume can be and they support live resize under NTFS/ReFS.

Can you migrate an LV from one PV to another?

Yes.

Can you do all kinds of creative striping, mirroring, and spanning?

Yes. Spanning is supported by default. Mirror (2-4 copies) and parity spaces can be selected at volume creation time. You can't layer then like you can in LVM though.

Overall the implementation is a lot like LVM (e.g. space metadata stored on the disk), so if a server fails with a mirrored/parity space you can just plug the disks into another machine and bring it up.


Coming in Server 2016 are:

Storage Spaces Direct: Storage spaces but between servers, basically LVM with remote PVs (this assumes you're running fast interconnects between servers).
Combine these using CSV and you've got something a bit like GPFS. That said, the first release is targeted at VM storage rather than standard files.

Storage Replica: Mirroring of volumes between servers (Basically DRBD).

[Kushan]

Not necessarily, comparing ntfs-3g with another FUSE&non-FUSE filesystem would be enough to find out how much overhead there is on Linux. Then we'd know which slowness is due to Windows instead of due to NTFS.

That wouldn't even remotely be a fair comparison. Comparing ntfs3g on linux with NTFS on Windows? You would glean nothing from that. Comparing ntfs-3g with a different filesystem on linux? Again, you would glean nothing other than the performance of ntfs-3g, which is not the same as the NTFS on Windows - completely different codebase.

You have no way of comparing NTFS performance on a Linux system, the best you can do is compare NTFS on Windows to another filesystem on Windows.

[Iksf]

I don't think anyone is suggesting that these tests give an accurate representation of day to day NTFS performance for Windows, obviously some bottleneck or work in progress there. I mean NTFS isn't great but its not HFS!

I... thought that Linux would blow it out of the water a bit. A bit disappointed.

Only so much impact the kernel has on performance. You're not going to get magical 500% performance wins running on another kernel unless something is wrong. The Windows kernel isn't bad by any stretch of imagination, its very competitive with Linux on speed. Even the BSD kernels are pretty competitive on speed and they have far less development resources.

[duby229]

That wouldn't even remotely be a fair comparison. Comparing ntfs3g on linux with NTFS on Windows? You would glean nothing from that. Comparing ntfs-3g with a different filesystem on linux? Again, you would glean nothing other than the performance of ntfs-3g, which is not the same as the NTFS on Windows - completely different codebase.

You have no way of comparing NTFS performance on a Linux system, the best you can do is compare NTFS on Windows to another filesystem on Windows.

Isn't comparing the different codebases the whole point? In your opinion then comparing anything with different code doesn't make sense? Amdgpu vs Catalyst? MS Office vs LibreOffice?

Determining the capability and performance of different implementations at the settings you use is kinda the whole point.

[MattFulkerson]

These benchmarks have me worried about the performance of Ubuntu 16.04 for numerical computations.

[gerddie]

Here found a way to run graphical applications and even entire DE, but not performance tested. Now you can test them.
P.S. Sorry, English is not the native language for me

Actually I would have been surprised if this wouldn't have worked. After all one can run applications remotely in Xming, all that is needed is a working network setup and the right display variable set.

[Kushan]

Isn't comparing the different codebases the whole point? In your opinion then comparing anything with different code doesn't make sense? Amdgpu vs Catalyst? MS Office vs LibreOffice?

Determining the capability and performance of different implementations at the settings you use is kinda the whole point.

No, I am saying Comparing the different code bases is simply not possible because one codebase only runs on Linux and one codebase only runs on Windows. That's not a fair test in the slightest, we still wouldn't know if it was NTFS that was slow or just Windows.

[duby229]

No, I am saying Comparing the different code bases is simply not possible because one codebase only runs on Linux and one codebase only runs on Windows. That's not a fair test in the slightest, we still wouldn't know if it was NTFS that was slow or just Windows.

I don't understand, filesystem formats are really just a layout of bits on a medium. The filesystem itself doesn't have any performance, it's just a layout. The driver implementation which reads and writes that layout is where performance is measured. Does it matter which driver is used or what OS the driver runs on?