The author of the benchmark here. I went over some comments and I'll try to tackle them here.
I'm pretty clear that GH runner based benchmark is far from perfect due to noisy neighbours etc. Thus every test first is running so called calibration... to reject completely unreliable VMs.
I'm fully aware that this can't completely fix the issue. Can limit it but not fix.
But as of now there are 593 runs recorded so average should still be quite meaningful.
Having that said I'm desperately trying to get REAL hardware to run that benchmark. With some successes ;)
But this new hardware has also so many disks that the plan is to try also more complex, tiered cache topologies. I'm working on it.
I'm happy to answer any other questions, sources of every piece of this benchmark are freely available and I'm not saying they are 100% correct. I'm open to improvements.
I went back and forth with Hetzner a couple times, I think we just got a bad machine :)
I've been saying it for months, but eventually I'm going to move the automated builds off the 48 core monster and we'll be able to use that for automated perf testing too. The machine we just got has spindles for EC perf testing, but the Hetzner monster has very high end enterprise ssdd.
> CI runs use loop devices on shared ephemeral VMs (one VM per filesystem): compare shapes and ratios, not absolute MB/s. Each job records a host-calibration anchor — see the table.
I think if you're not using baremetal for such tests, it's likely that the results are simply not comparable at all? What if another tenant is also using the disk?
It's a fair point but it's also possible the person running the tests has a dedicated test hypervisor for this , so that different configurations of filesystems and VMs can be created and destroyed quickly in an automated manner.
If it's something as simple as a KVM hypervisor that only runs 1 test VM at a time (with no other load from anything else other than the basic systemd daemons, ssh daemon etc running on the hypervisor), the results could be very close to bare metal.
I can see it being very time consuming and annoying to do repeated manual bare metal OS installs and new partitioning/filesystem creation for such a large variety of tests.
The author does also say that performance isn't really the main thing but rather, data integrity:
> I can see it being very time consuming and annoying to do repeated manual bare metal OS installs.
Well don't do that then. There's lots of other options. Probably the simplest is a single bare metal install on a simple filesystem on one device. run the filesystems under test on other storage dedicated to testing.
You could also boot into a network install and use local storage exclusively for testing.
Yes exactly. The issueThe epherrality of the VMs isn't an issue, it's the _shared_ part that's the concern here. Going by the fact that the kernel is listed as "kernel 7.0.0-1012-azure" I feel like it's a fair risk that there may have been noisy neighbours.
In this case, given that the author's own disclaimer (above) already disclaims the numeric readings, I'm not sure how it's possible to make any inference on "shapes and ratios" derived from the numeric readings.
> Every push/2-hourly cron builds each filesystem across 4 loop devices backed by sparse files, runs the suite, and publishes a results table in the job summary plus JSON artifacts.
I get that real hardware costs (author mentions EUR 70 a month for a suitable server), but without at least a baseline snapshot comparison run between real hardware, both SSD and HDD, and the sparse file-backed loop devices, it's hard to take much away from this.
Sadly the AI apocalypse isn't making stuff like this easy to do as a hobby.
What is md-raid10 doing that is so much worse than lvm-raid10? In terms of "I/O" and "responsiveness." It's not really obvious to me from either the linked page or https://github.com/fenio/modern-fs-benchmark . In principle they should be similar?
I think the reviews should also include the social aspect of these filesystems...
There is and have been many promising and exciting FS to replace the old boring ones, but for storage you not only want to avoid technical issues but also maintainer(s) drama...
To answer the original question, most people who care about their filesystem at all care about its stability. Not just "does it work now" but also "will it work and improve over time". Infighting puts the future at risk.
That might be the best thing happened to the project since now development can happen at its own pace without the clicky bait influencers.
In fact they delivered the erasure coding for parity raid back in march this year.
The thing is that as soon as you seriously give a chance to Bcachefs you see how good it is. I can only tell you that mixing different device tiers and having a per-file/directory replication setting is a god send specially in these times where storage costs more than gold.
So two filesystems that are essentially shunned from the Linux kernel and permanent second-class citizens, and one that was removed from Red Hat and has a questionable history of reliability. Oh boy which do I choose?
@farlight assuming that you're the creator do you think you'd be able to rework the HTML/CSS? I'm sure you've got good data but speaking on behalf of my eyeballs, the results page is... hard to read!
> 2G of random garbage is written directly onto one member device (behind the filesystem's back, offset 1G — python injector; uutils dd mis-seeks on dm devices), caches dropped, then a full scrub: btrfs scrub -B, zpool scrub + wait, bcachefs scrub, md/lvm sync-action 'check' (which can only COUNT mismatches — no checksums to know which copy is right).
I'm not sure that nuking 2G of the underlying block device is a recoverable error on any filesystem that I'm aware of? Can you confirm if any ofthe filesystems really came out of the other side in a usable state after scrubbing?
-----
> Trivial-op p99, idle (ms) # A trivial operation — one 4k write + fsync every 200ms (like a shell appending history or an editor updating its swap file) — run alone for 10s. p99 of the fsync completion
In fact, if it's OK for me to ask, are any of the metrics tht you used standard industry metrics? It looks like several of the tests are bypassing the kernel's page cache? -- which I worry may fall into the trap of "I modified the system to be unrepresentative of reality and then tested it".
----
> kernel 7.0.0-1012-azure
Can you confirm if you tested on a bare metal machine? were you the only tenant?
Having that said I'm desperately trying to get REAL hardware to run that benchmark. With some successes ;)
Few months ago I got Hetzner machine from Kent Overstreet and I was able to finish 3 runs before machine died... Results: https://bartosz.fenski.pl/modern-fs-benchmark/real-hw/
Currently I've got even more interesting machine with tons of disks and I'm running new set of benchmarks but it's really in its initial stage.
https://bartosz.fenski.pl/modern-fs-benchmark/sas-hdd/ 2nd run in progress... one run on REAL hardware takes much more time than on GH runner so it's slow.
But this new hardware has also so many disks that the plan is to try also more complex, tiered cache topologies. I'm working on it.
I'm happy to answer any other questions, sources of every piece of this benchmark are freely available and I'm not saying they are 100% correct. I'm open to improvements.
I've been saying it for months, but eventually I'm going to move the automated builds off the 48 core monster and we'll be able to use that for automated perf testing too. The machine we just got has spindles for EC perf testing, but the Hetzner monster has very high end enterprise ssdd.
Also, just got done with the Rust for Linux conference, still not home but here's slides that still need reformatting: https://evilpiepirate.org/~kent/Kangrejos-2026-bcachefs.pdf
I think if you're not using baremetal for such tests, it's likely that the results are simply not comparable at all? What if another tenant is also using the disk?
If it's something as simple as a KVM hypervisor that only runs 1 test VM at a time (with no other load from anything else other than the basic systemd daemons, ssh daemon etc running on the hypervisor), the results could be very close to bare metal.
I can see it being very time consuming and annoying to do repeated manual bare metal OS installs and new partitioning/filesystem creation for such a large variety of tests.
The author does also say that performance isn't really the main thing but rather, data integrity:
https://github.com/fenio/modern-fs-benchmark
Well don't do that then. There's lots of other options. Probably the simplest is a single bare metal install on a simple filesystem on one device. run the filesystems under test on other storage dedicated to testing.
You could also boot into a network install and use local storage exclusively for testing.
In this case, given that the author's own disclaimer (above) already disclaims the numeric readings, I'm not sure how it's possible to make any inference on "shapes and ratios" derived from the numeric readings.
I get that real hardware costs (author mentions EUR 70 a month for a suitable server), but without at least a baseline snapshot comparison run between real hardware, both SSD and HDD, and the sparse file-backed loop devices, it's hard to take much away from this.
Sadly the AI apocalypse isn't making stuff like this easy to do as a hobby.
There is and have been many promising and exciting FS to replace the old boring ones, but for storage you not only want to avoid technical issues but also maintainer(s) drama...
The community infighting has sucked, but that's a thing that matters primarily for maintainers.
I think most users just want something that works.
To answer the original question, most people who care about their filesystem at all care about its stability. Not just "does it work now" but also "will it work and improve over time". Infighting puts the future at risk.
I do have a lot more pull requests to merge than I did before. I don't know if you want to count "Kent isn't reviewing Pars fast enough" as drama :)
In fact they delivered the erasure coding for parity raid back in march this year.
The thing is that as soon as you seriously give a chance to Bcachefs you see how good it is. I can only tell you that mixing different device tiers and having a per-file/directory replication setting is a god send specially in these times where storage costs more than gold.
I'm saying ZFS on another OS.
I'm not sure that nuking 2G of the underlying block device is a recoverable error on any filesystem that I'm aware of? Can you confirm if any ofthe filesystems really came out of the other side in a usable state after scrubbing?
-----
> Trivial-op p99, idle (ms) # A trivial operation — one 4k write + fsync every 200ms (like a shell appending history or an editor updating its swap file) — run alone for 10s. p99 of the fsync completion
In fact, if it's OK for me to ask, are any of the metrics tht you used standard industry metrics? It looks like several of the tests are bypassing the kernel's page cache? -- which I worry may fall into the trap of "I modified the system to be unrepresentative of reality and then tested it".
----
> kernel 7.0.0-1012-azure
Can you confirm if you tested on a bare metal machine? were you the only tenant?