Btrfs RAID 0/1 Benchmarks On The Linux 4.1 Kernel

[ormaaj]

Looking forward to the RAID 10 tests as I suspect that's what almost everyone using BTRFS RAID today is using including myself (6x 2Tb SATA drives).

[Ardje]

Theoretically RAID1 can be as fast as all disks combined for sequential reads as long as the RAID controller supports proper load balancing.

That's disk based raid1, btrfs raid1 is different: btrfs raid 1 on data means it has a copy of your file on at least one other disk. It uses the checksums to see which one is correct.
In that respect, btrfs is so much better than block based raid1 because btrfs knows which copy is not corrupt.
Btrfs raid works on "file level" (extent level is more like it I guess), so on a 2 disk raid 1 the layout on one disk can be totally different from the other disk.
So as soon as BTRFS is really trustworthy stable, I will use it on bare disks, and scrap my md-raid. But for now md-raid is proven technology and fast too.

[kobblestown]

And since when does anyone use four drives for raid 1?

I do. But with Btrfs it's a bit different from what you would expect. It doesn't keep four copies of the same data. It still only keeps two copies on different disks. The advantage is that you can have, say, three 1TB disks and one 3TB disk and you get 3TB mirrored array.

Did he mean to test 0+1?

I'd like to see such tests too. But I doubt that's what Michael did because it should have performed quite a bit better than two disk RAID1.

[terrywang]

Recently I've built a NAS (MicroServer N40L, running Fedora 21) for my colleague, with 5 disks. 4*2TB WD Green form a Btrfs RAID5.

I've tired the write to Btrfs RAID5, the write speed (using both 4K and 1M block size) ranged from 31x MB/s ~ 330 MB/s the guy was pretty happy about the result.

Screenshot -> https://flic.kr/p/sbZEVx

[rdnetto]

I'd really like to see some benchmarks analyzing the performance of btrfs with snapshots (since that's why most people use it). Something like:
1. write data
2. snapshot
3. read & write, modifying ~5% of the data
4. repeat from 2.

My anecdotal experience suggests that it slows down massively for the common use case, but it'd be nice to see some verification of this.

[toyotabedzrock]

Theoretically RAID1 can be as fast as all disks combined for sequential reads as long as the RAID controller supports proper load balancing. I'm not sure about how btrfs implements load balancing. It rarely happens though because most reads aren't very sequential. You'll only get maximum performance when reading large sequential files and most linux systems are comprised of thousands of very tiny files where access latency becomes the bottleneck.

It should speed up any reads, after the reads are reorganized for head position it should just send batches to both.
And since when does anyone use four drives for raid 1? Did he mean to test 0+1?

[duby229]

Curiously, does compressing the small-files sections (like the system directories), particularly if they're text, help reduce the this access? I know it used to in to a degree on the old spinning discs. But nowadays, all this newandangled stuff...

Well, RAID on SSDs has entirely different characteristics, Access latency is smaller problem.

Speaking about compressing small files, I keep a squashfs filesystem mounted just for that purpose. It gets synced every night so the next days filesystem is last nights image. It really does help improve performance when accessing many small files.

[stiiixy]

Theoretically RAID1 can be as fast as all disks combined for sequential reads as long as the RAID controller supports proper load balancing. I'm not sure about how btrfs implements load balancing. It rarely happens though because most reads aren't very sequential. You'll only get maximum performance when reading large sequential files and most linux systems are comprised of thousands of very tiny files where access latency becomes the bottleneck.

Curiously, does compressing the small-files sections (like the system directories), particularly if they're text, help reduce the this access? I know it used to in to a degree on the old spinning discs. But nowadays, all this newandangled stuff...

[Fry-kun]

Nice benchmarks!
Mind including the "single JBOD" config in the tests, too?

[duby229]

Theoretically RAID1 can be as fast as all disks combined for sequential reads as long as the RAID controller supports proper load balancing. I'm not sure about how btrfs implements load balancing. It rarely happens though because most reads aren't very sequential. You'll only get maximum performance when reading large sequential files and most linux systems are comprised of thousands of very tiny files where access latency becomes the bottleneck.