Btrfs Working On RAID1 Round-Robin Read Balancing

[phoronix]

Phoronix: Btrfs Working On RAID1 Round-Robin Read Balancing

It's been a while since there has been anything new to report on the Btrfs file-system's built-in RAID functionality but that is changing with RAID1 round-robin read balancing...

https://www.phoronix.com/news/Btrfs-RAID1-Round-Robin-Coming

[bezirg]

Wait, how did RAID1 work previously in btrfs?

[npwx]

That seems like such a basic feature, I'm really surprised they did not have that yet.

[Ananace]

Wait, how did RAID1 work previously in btrfs?

It wrote the data in duplicated stripes across the disks, and then it would read from one of the stripes, selected based on the pid of the reading process.
With this change it can select stripes based on round-robin for reads instead.

[npwx]

It wrote the data in duplicated stripes across the disks, and then it would read from one of the stripes, selected based on the pid of the reading process.
With this change it can select stripes based on round-robin for reads instead.

So it left all the read performance benefits of RAID1 unused under the table. Given the age of btrfs, this leaves me baffled. I can see plenty of micro optimizations over time, yet this big one took what, 10, 15 years?

[woddy]

So it left all the read performance benefits of RAID1 unused under the table. Given the age of btrfs, this leaves me baffled. I can see plenty of micro optimizations over time, yet this big one took what, 10, 15 years?

...or maybe it wasn't a priority for them.

[microcode]

So it left all the read performance benefits of RAID1 unused under the table. Given the age of btrfs, this leaves me baffled. I can see plenty of micro optimizations over time, yet this big one took what, 10, 15 years?

Rarely matters. They call it raid1 in the config for familiarity but it really is just mirroring. It is also a more complex thing on btrfs than on block device RAID (where the block addresses are the same for everything).

[waxhead]

So it left all the read performance benefits of RAID1 unused under the table. Given the age of btrfs, this leaves me baffled. I can see plenty of micro optimizations over time, yet this big one took what, 10, 15 years?

No. The pid based approach actually works quite well when there are many processes accessing the storage devices. This should work well for single processes depending on your hardware. There are far more interesting patches in the pipeline than the round robin method as well which takes into account how busy the devices are. Personality i ponder if not something like the principles of the EEVDF might be a good choice here as well.

[gotar]

The pid based approach actually works quite well when there are many processes accessing the storage devices.

At least at a time when RAID1 was limited to 2 devices (no 3+way mirroring) they used PID parity. So more processes could be reading from the same device ignoring entirely second one (of course with increasing number of processes the probability of this decreases, but it's obviously suboptimal).

And some people still believe this is production ready fs...

[pWe00Iri3e7Z9lHOX2Qx]

So it left all the read performance benefits of RAID1 unused under the table. Given the age of btrfs, this leaves me baffled. I can see plenty of micro optimizations over time, yet this big one took what, 10, 15 years?

Btrfs should never have used the term "RAID", because those levels actually mean something to most people. The write or in this case read strategies are often different than what you would assume based on the RAID level. Like the write strategy for RAID 10 reordering disks meaning you are basically guaranteed to lose data when any 2nd disk fails, compared to a traditional RAID 10 where you are only screwed if the 2nd failed disk is from the same mirror. So you go from a 33% chance of losing data in a traditional four disk RAID 10 second failed disk scenario to a 100% chance with Btrfs RAID 10. And this isn't obvious from the documentation. Fun right?

Btrfs is designed to be as flexible as possible with maximizing storage / mixing disks / changing "RAID" levels. That can be great if that's the most important thing to you. But there are a slew of reasons that many people put up with the PITA that comes with out of tree modules and use ZFS.