From: Szabolcs Szakacsits <firstname.lastname@example.org>
To: Linus Torvalds <email@example.com>
Cc: Theodore Ts'o <firstname.lastname@example.org>,
Matthew Wilcox <email@example.com>,
"Leonidas P. Papadakos" <firstname.lastname@example.org>,
Konstantin Komarov <email@example.com>,
<firstname.lastname@example.org>, "Darrick J. Wong" <email@example.com>,
Greg Kroah-Hartman <firstname.lastname@example.org>,
Hans de Goede <email@example.com>,
Linux Kernel Mailing List <firstname.lastname@example.org>,
Al Viro <email@example.com>,
Miklos Szeredi <firstname.lastname@example.org>
Subject: Re: NTFS testing (was: [GIT PULL] vboxsf fixes for 5.14-1
Date: Fri, 3 Sep 2021 01:09:40 +0300 [thread overview]
Message-ID: <alpine.DEB.email@example.com> (raw)
On Tue, 3 Aug 2021, Linus Torvalds wrote:
> On Tue, Aug 3, 2021 at 5:04 PM Theodore Ts'o <firstname.lastname@example.org> wrote:
> > Whenver I've ever needed to access ntfs files, I've always used the
> > ntfs-3g FUSE package.
> The user-space FUSE thing does indeed work reasonably well.
> It performs horribly badly if you care about things like that, though.
> In fact, your own numbers kind of show that:
> ntfs/default: 670 tests, 55 failures, 211 skipped, 34783 seconds
> ntfs3/default: 664 tests, 67 failures, 206 skipped, 8106 seconds
> and that's kind of the point of ntfs3.
In all fairness, the generic/405 test case completely distorted the overall
timing in favour of ntfs3. Neither driver was involved in that test case.
Generic/405 test is mkfs against a 1 TB thin provision device which has 1
MB backing size. mkfs should return an error after it hits EIO. The test
case configuration was not correct for mkntfs on behalf of ntfs-3g because
it missed the --fast format option, so mkntfs tried to fill the 1 TB device
with zeros, apparently on Google Cloud Platform, for almost 8 hours. This
had absolutely nothing to do with ntfs-3g performance, it was a pure mkfs
Meanwhile the test case ran in 1 second on behalf of ntfs3 because its mkfs
was not found, so the test could not be run. (And the test case got
incorrectly categorized as success because it interpreted the "command not
found" error as a success.)
If this mkntfs test case is ignored, as it should be, then ntfs-3g's
runtime was (34783 - 28396) = 6387 versus ntfs3's (8106 - 1) = 8105
seconds, i.e. the user space ntfs-3g was about 21% faster overall than the
kernel space ntfs3.
Does this mean ntfs-3g is faster than ntfs3? Of course not. Fstests is not
a benchmark. What we know for sure is, the unknowingly configured, untuned
versions of the software gave different times for different workloads.
File system performance is a fairly complex topic. Ntfs-3g always aimed for
stability, features, interoperability and portability, not for best
possible performance. There seems to be some misconceptions,
misinterpretations, inefficient configuration and mount options (e.g.
missing big_writes, kernel_cache, etc). Unfortunately we did our part too
to end up here. We will set better performance defaults in future releases.
User space drivers can have major disadvantages for certain workloads
however how relevant are those for NTFS users? Most people use NTFS for
file transfers in which case ntfs-3g read and write speed is about 15-20%
less compared to ext4. For example in some quick tests ext4 read was
3.4 GB/s versus ntfs-3g 2.8 GB/s, and write was 1.3 GB/s versus 1.1 GB/s.
Additionally there are still several technical solutions which could be
implemented to improve all kinds of user space driver performance
But again, we always prefer data integrity over performance. And NTFS can
be quite tricky with the ever changing on-disk corner cases. Does anybody
still remember when Windows 2000 changed the NTFS on-disk format which
massively started to trash users' data?
Please don't get me wrong, I'm not saying this is the way to go (who would
be so crazy to write anything like that on the linux-kernel list?) I'm just
saying this is the way we chose and support. We welcome the recent interest
in NTFS after working on it for 20 years.
These are from Ted's logs which he shared earlier. It's much appreciated,
it was highly useful. Personally I also thought the very poor ntfs-3g
timing was due to bad configuration and/or mount options instead of an
irrelevant test case. (Btw, the driver configuration and mount options were
indeed not right, e.g. ACL, permission, etc related cases failed which
could have pass.)
$ egrep ^generic/405 results-ntfs*/runtests.log
results-ntfs/runtests.log:generic/405 [21:47:08] [05:40:24] 28396s
results-ntfs3/runtests.log:generic/405 [12:12:09] [12:12:10] 1s
$ cat results-ntfs/ntfs/results-default/generic/405.full
Cluster size has been automatically set to 4096 bytes.
Initializing device with zeroes: 100% - Done.
Creating NTFS volume structures.
Failed to sync device /dev/mapper/thin-vol: Input/output error
Syncing device. FAILED
$ cat results-ntfs3/ntfs3/results-default/generic/405.full
mkfs: failed to execute mkfs.ntfs3: No such file or directory
next prev parent reply other threads:[~2021-09-02 22:49 UTC|newest]
Thread overview: 42+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-07-13 10:45 [GIT PULL] vboxsf fixes for 5.14-1 Hans de Goede
2021-07-13 19:15 ` Linus Torvalds
2021-07-13 20:14 ` Al Viro
2021-07-13 20:18 ` Al Viro
2021-07-13 20:24 ` Randy Dunlap
2021-07-13 20:32 ` Al Viro
2021-07-13 21:43 ` Randy Dunlap
2021-07-14 10:50 ` Rafał Miłecki
2021-07-14 14:13 ` Christoph Hellwig
2021-07-14 14:51 ` Greg KH
2021-07-14 15:59 ` Rafał Miłecki
2021-07-14 16:05 ` Matthew Wilcox
2021-07-14 16:18 ` Rafał Miłecki
2021-07-15 21:50 ` Neal Gompa
2021-07-16 11:46 ` Leonidas P. Papadakos
2021-07-16 18:07 ` Linus Torvalds
2021-07-30 15:55 ` Konstantin Komarov
2021-07-30 17:23 ` Paragon NTFSv3 (was Re: [GIT PULL] vboxsf fixes for 5.14-1) Linus Torvalds
2021-08-13 16:11 ` Konstantin Komarov
2021-08-15 20:32 ` Stephen Rothwell
2021-08-16 3:00 ` Kari Argillander
2021-09-02 21:55 ` Linus Torvalds
2021-08-03 22:48 ` [GIT PULL] vboxsf fixes for 5.14-1 Theodore Ts'o
2021-08-03 23:44 ` Matthew Wilcox
2021-08-04 0:04 ` Theodore Ts'o
2021-08-04 0:10 ` Linus Torvalds
2021-08-04 0:49 ` Theodore Ts'o
2021-08-04 1:03 ` Darrick J. Wong
2021-08-04 6:38 ` Kari Argillander
2021-08-04 16:30 ` Theodore Ts'o
2021-08-05 15:48 ` Konstantin Komarov
2021-08-10 7:02 ` Darrick J. Wong
2021-09-02 22:09 ` Szabolcs Szakacsits [this message]
2021-09-03 17:48 ` NTFS testing (was: " Eric Biggers
2021-09-03 21:17 ` Szabolcs Szakacsits
2021-07-17 16:47 ` Pali Rohár
2021-07-14 16:13 ` Darrick J. Wong
2021-07-14 16:18 ` Christoph Hellwig
2021-07-14 16:38 ` Gao Xiang
2021-07-14 20:03 ` Eric W. Biederman
2021-07-15 22:14 ` Darrick J. Wong
2021-07-13 19:17 ` pr-tracker-bot
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).