DragonFly's HAMMER2 File-System Picks Up More Performance Optimizations

[starshipeleven]

But perhaps HAMMER2 can offer other things besides performance such as those features offered by the likes of Btrfs and ZFS.

some yes, some no https://gitweb.dragonflybsd.org/drag...hammer2/DESIGN

And it has similar issues like "no free space on device" when it's not really true https://garyshood.com/hammer2-space/ which is a hard thing to deal with for CoW filesystems in general.

[uid313]

Michael did benchmark it some months ago https://www.phoronix.com/scan.php?pa...x-Initial-Data

it seems to do some shenanigans when doing a fsync (i.e. it's flat-out lying about the time it takes to write to disk, which is roughly the same on all filesystems as the disk is the same), apart from that it's significantly slower.

He will probably do it again eventually, and will see how it fares after these improvements.

I have some strong suspicions that it can't do significantly better than ext4 or xfs, there is only so much you can do with a traditional filesystem.

I have a feeling you're right about the performance. Ext4 and XFS has been around for a long while, are extensively used and are probably pretty tuned and performant.

But perhaps HAMMER2 can offer other things besides performance such as those features offered by the likes of Btrfs and ZFS.
Online fsck, online resizing, transparent encryption, transparent compression, etc. Maybe RAID-like things and volume management, but maybe there is LVM for that.

[starshipeleven]

But even without the cluster functionality, maybe it is good as a normal non-cluster file system?
I don't know how it is compared to ext4, reiser4, Btrfs and ZFS.

Michael did benchmark it some months ago https://www.phoronix.com/scan.php?pa...x-Initial-Data

it seems to do some shenanigans when doing a fsync (i.e. it's flat-out lying about the time it takes to write to disk, which is roughly the same on all filesystems as the disk is the same), apart from that it's significantly slower.

He will probably do it again eventually, and will see how it fares after these improvements.

I have some strong suspicions that it can't do significantly better than ext4 or xfs, there is only so much you can do with a traditional filesystem.

[uid313]

It could be interesting if HAMMER2 was actually finished and not in a sad sorry state like now (i.e. it's supposed to be a cluster filesystem and it works only on the local machine for now), but even if it was finished, Linux has like 3-4 production-grade cluster filesystems already, so whatever.

But even without the cluster functionality, maybe it is good as a normal non-cluster file system?
I don't know how it is compared to ext4, reiser4, Btrfs and ZFS.

[starshipeleven]

Um, what I said.

not clear enough, I guess.

[Bsdisbetter]

FYI: it's unlikely to get past core because it's going to require significant change everywhere.

Also I don't see other bsd projects caring about this, given their goals.

Um, what I said.

[starshipeleven]

Maybe it would be nice if HAMMER2 was available for Linux.
But maybe it would be difficult to port it from DragonFly BSD to Linux.

It could be interesting if HAMMER2 was actually finished and not in a sad sorry state like now (i.e. it's supposed to be a cluster filesystem and it works only on the local machine for now), but even if it was finished, Linux has like 3-4 production-grade cluster filesystems already, so whatever.

[starshipeleven]

Nothing's impossible in that regard, it is do-able, it's just that you're unlikely to get it past core.

FYI: it's unlikely to get past core because it's going to require significant change everywhere.

Also I don't see other bsd projects caring about this, given their goals.

[aht0]

Btw, Hammer and Hammer 2 are different animals.

[Bsdisbetter]

Well at the moment it's apparently too difficult to port it to even other BSD:s https://wiki.freebsd.org/PortingHAMMERFS

Nothing's impossible in that regard, it is do-able, it's just that you're unlikely to get it past core. They're an immovable object sometimes, however, their reasoning on this is sound.

Previously, I think it was tried with openbsd and I recall looking at the code changes and thinking, yowee, that's a lot of work for integrating a file system when freebsd has zfs which is both a file system and volume manager.

I can understand netbsd or openbsd attempting to use it, not freebsd.

A quick check finds this benchmark: https://www.phoronix.com/scan.php?pa...x-Initial-Data

Hmm...