Corsair USB 3.0 Flash Voyager Drives: EXT4 vs. NTFS vs. Btrfs vs. F2FS

[willmore]

Seems there was typo? 1GiB USB 3.0 drive is a nice joke.

I mentioned that in comment #10 already.

[gilboa]

I was talking of desktop use-case mostly (most common place where you will find NTFS all over the place). There it's fast as anything else. Not like 400% slower than ext4 on any test like shown here.

I guess you're right, though it does depends on the benchmark.
E.g. Large project compilation may trigger the same issues w/ NTFS (Slow file creation, slow access to huge-directories) as we witnessed.

- Gilboa

[starshipeleven]

@starshipeleven,

Performance wise (especially in large-storage-cases), NTFS is a rotten file system, even on Windwos 2K8/12.
We've developed a big data software that uses POSIX (Linux) and Win32 (Windows) to create billions of small (<64KB) compressed files.
No matter what we do, and that includes using low-level internal API (NtCreateFile, etc) NTFS performance remains 80% (if not more) slower than ext4.
Far worse, once you reach more than 1000 files per sub-directory, file creation (via NtCreateFile) becomes 100 (!) times slower than POSIX creat.

Granted, once you have the file handles active, read / write performance are more or less the same.

- Gilboa

I was talking of desktop use-case mostly (most common place where you will find NTFS all over the place). There it's fast as anything else. Not like 400% slower than ext4 on any test like shown here.

[gilboa]

This isn't NTFS from windows, but NTFS from linux. Linux NTFS drivers are crappy because various obvious reasons. NTFS from windows is pretty good.

exFAT from linux uses the same FUSE infrastructure of NTF on linux, so it will have crap performance, and might be also as unreliable.

@starshipeleven,

Performance wise (especially in large-storage-cases), NTFS is a rotten file system, even on Windwos 2K8/12.
We've developed a big data software that uses POSIX (Linux) and Win32 (Windows) to create billions of small (<64KB) compressed files.
No matter what we do, and that includes using low-level internal API (NtCreateFile, etc) NTFS performance remains 80% (if not more) slower than ext4.
Far worse, once you reach more than 1000 files per sub-directory, file creation (via NtCreateFile) becomes 100 (!) times slower than POSIX creat.

Granted, once you have the file handles active, read / write performance are more or less the same.

- Gilboa

[ryao]

Phoronix: Corsair USB 3.0 Flash Voyager Drives: EXT4 vs. NTFS vs. Btrfs vs. F2FS

With having some new Corsair USB 3.0 Flash Voyager flash drives around, I decided to run some fresh Linux file-system benchmarks on them to see how various file-systems are performing on low-cost USB flash drives.

http://www.phoronix.com/vr.php?view=23096

Michael was right to remark about the FIO results being wrong, but he was wrong to claim it is fsync related on a read only benchmark. He should have claimed that it is cache related. Doing O_DIRECT in FIO evidently has no effect on how the FUSE driver opens the block device.

That means that the FUSE driver is reading out of the kernel buffer cache. This could be confirmed with strace. iostat should also show 0 IO requests being sent to the block device.

That also means that Michael failed to scale the benchmark's dataset to exceed system memory. Doing twice system memory as Brendan Gregg does would have eliminated the buffer cache as a confounding variable and produced a reliable FIO result (assuming no other confounding variables like sector misalignment).

[SystemCrasher]

This Corsair USB 3.0 Flash Voyager 1GB drive

Seems there was typo? 1GiB USB 3.0 drive is a nice joke. And, hmm, yeah, FUSE suxx when it comes to speed. NTFS should be last resort option for Linux users.

[Passso]

This isn't NTFS from windows, but NTFS from linux. Linux NTFS drivers are crappy because various obvious reasons. NTFS from windows is pretty good.

exFAT from linux uses the same FUSE infrastructure of NTF on linux, so it will have crap performance, and might be also as unreliable.

Must agree with you on this point, without adding the "real" NTFS on Windows for comparison in this test, talking about NTFS is not fair.

[drSeehas]

Only FAT, vFAT, exFAT support the uid, gid, umask option.

VFAT is only an extension on top of FAT.

[dragonn]

And ntfs-3g.

FUSE -.- I want native file system, now I am using exfat with native kernel driver but it is not so fast like ext4 :/.

[starshipeleven]

Is there an in-kernel MS-friendly filesystem that runs with good performance I can use in a linux distro? I dual-boot at the moment and reserved a large storage partition for some games and scratch discing etc, and yeah, NTFS is rather slow when in linux. There's also the fun bit of having to restart in Windows if it went to sleep while in Windows otherwise the partition is marked unclean and drops to console for recovery. I believe that may well be caused by swap on that partition (hiber is actually turned off)

Short answer: no.

Longer answer: probably FAT32, but can also happen that you write on a FAT32 drive from linux and then Windows wil NOT RECOGNIZE THE PARTITION ANYMORE, while on linux it is still perfectly fine and readable and fs checks don't find any error.

Possible Future Answer: maybe F2FS, it was planned to have a Windows version too.