FUSE Patch Allows Non-Extending, Parallel Direct Writes For Better Performance

[phoronix]

Phoronix: FUSE Patch Allows Non-Extending, Parallel Direct Writes For Better Performance

For those making use of FUSE user-space file-system capabilities, developer Dharmendra Singh has posted a patch to allow for non-extending parallel direct writes. In turn for threaded write scenarios this can mean a huge boost to performance...

https://www.phoronix.com/scan.php?page=news_item&px=FUSE-Parallel-Direct-Writes

[aorth]

Could be huge for GlusterFS?

[Danny3]

Wonderful!

[caligula]

Wonderful!

Now that exfat and ntfs are both provided by efficient in-kernel drivers, what are the most interesting fuse modules to use?

[Rallos Zek]

Now that exfat and ntfs are both provided by efficient in-kernel drivers, what are the most interesting fuse modules to use?

NTFS-3g so far is proven to better then the new in-kernel ntfs driver in terms of stability and being more bug-free.

[Dark-Show]

I wonder if this would fix that "torrent corruption" bug (that people talk about on here) with NTFS-3g?

[yump]

Now that exfat and ntfs are both provided by efficient in-kernel drivers, what are the most interesting fuse modules to use?

Perhaps sshfs? Kde connect uses it for sharing the phone a filesystem, I think.

The main use I can think of for parallel non-extending writes to the same file is BitTorrent though. So maybe this will help mergerfs.

[uid313]

So how fast is FUSE now compared to native in-kernel drivers?
Could ext4, Btrfs and ZFS moved out of the kernel and into FUSE?
Could we move out all the file system drivers from the kernel and run all file systems in FUSE?
Like a microkernel architecture.
Or is FUSE still way too slow even after this performance improvement?
Can FUSE performance be further improved? Can it reach native in-kernel speeds?

[Weasel]

Can it reach native in-kernel speeds?

No.

5 chars.

[daleto]

So how fast is FUSE now compared to native in-kernel drivers?
Could ext4, Btrfs and ZFS moved out of the kernel and into FUSE?
Could we move out all the file system drivers from the kernel and run all file systems in FUSE?
Like a microkernel architecture.
Or is FUSE still way too slow even after this performance improvement?
Can FUSE performance be further improved? Can it reach native in-kernel speeds?

This is a valid and interesting question
I/O in the way how it is design today usin POSIX comes with a cost as it uses CPU resources more than necessarily needed

L4se (which is a Micro kernel design) consider bildning a new OS as they see the real limitations of POSIX

I think that performance can absolutely be better in a fuse based filesystem than in a kernel, even though there are context switches between the tasks

Why better?
Because the kernel cuts all I/O to 4K segmentets, puts the requests on the I/O elevator which tried to glue it together again jag
I have seen applications using 256K requests where the elevator can't glue it together again, ended up with 80K
Moved that to fuse had plain 256K requests