XFS RAID0 Benchmarks Across Twenty SSDs vs. EXT4 & Btrfs On Ubuntu Linux

[phoronix]

Phoronix: XFS RAID0 Benchmarks Across Twenty SSDs vs. EXT4 & Btrfs On Ubuntu Linux

Earlier this month were the FreeBSD ZFS vs. Linux EXT4/Btrfs RAID With Twenty SSDs. Besides interest in seeing ZOL tests (they're already planned upon the ZFS On Linux 0.8 release), there was also some interest by readers in seeing some XFS RAID tests side-by-side. Here are some of those XFS RAID benchmarks up against Btrfs and EXT4 from Ubuntu Linux...

http://www.phoronix.com/scan.php?page=news_item&px=XFS-RAID0-Twenty-SSDs

[dimko]

It sucks to be unemployed, so i wont be buying any of those SSD any time soon, but as far as I can see, no point in using anything but EXT4 at this time.

[caligula]

It sucks to be unemployed, so i wont be buying any of those SSD any time soon, but as far as I can see, no point in using anything but EXT4 at this time.

64 GB SSDs are pretty cheap now. Around $20. You'll get 500 MB/s+ throughput and 100k+ IOPS.

[starshipeleven]

It sucks to be unemployed, so i wont be buying any of those SSD any time soon, but as far as I can see, no point in using anything but EXT4 at this time.

In actual day-to-day usage there isn't THAT much of a difference between low end or small capacity and high end / high capacity SSDs, as long as you are buying good brands and not random chinese crap.

The jump from mechanical to SSD is the only thing you will actually feel. Especially on a laptop.

[polarathene]

I've not used RAID before, though I'm vaguely familiar with how it works and the different kinds. RAID0 is for speed with 2+ disks right? How were the results here achieving sometimes 3x or more performance of a single disk?

Oh... nevermind, the RAID0 was 20x not 2x 860 evos.. In that case why is the perf so slow, what is the benefit to RAID0 with that many disks?

> A RAID 0 array of n drives provides data read and write transfer rates up to n times as high as the individual drive rates

That suggests the perf should be much better? From what I understand if all drive are the same capacity, then it accumulates the total capacity available rather than limiting capacity to that of a single disk... so is it more to do with that? Data is read/written in a way that it can spread of disks to maximize the I/O, but without enough data to saturate the I/O(which I'd have thought some benchmarks for this catered for), or the disk controllers bus bandwidth being a bottleneck, the perf caps, as it can't utilize the extra I/O perf that many disks would offer?

no point in using anything but EXT4 at this time.

How do you figure that? XFS looked like it did pretty well in these benchmarks, better than EXT4?

[tului]

Given that performance isn't anywhere near 20X I'd think RAID 10 would be better. At least some tiny protection

[starshipeleven]

what is the benefit to RAID0 with that many disks?

Running more parallel processes that access the disks without bottlenecking them.

From what I understand if all drive are the same capacity, then it accumulates the total capacity available rather than limiting capacity to that of a single disk... so is it more to do with that? Data is read/written in a way that it can spread of disks to maximize the I/O, but without enough data to saturate the I/O(which I'd have thought some benchmarks for this catered for), or the disk controllers bus bandwidth being a bottleneck, the perf caps, as it can't utilize the extra I/O perf that many disks would offer?

RAID0 divides the data into chunks called "stripes" that are subsequently written on the drives. The size of these chunks is the "stripe size".

If your read/writing activity fill only 3 stripes at a time, then you get only 3 times the performance of a single drive as you are actually using only 3 drives. This is a oversimplification for the sake of understanding the behaviour.

RAID stripe size and other settings are usually configured for the specific type of workload they will face, there is no "best configuration" as the workload is different.

I'd guess that such an array could deal with multiple benchmarks running at the same time and provide them the same bandwith as shown up to a certain point where it is saturated, while running multiple benchmarks on a single disk system would immediately divide the performance by the number of benchmarks running at the same time.

[xiando]

In actual day-to-day usage there isn't THAT much of a difference between low end or small capacity and high end / high capacity SSDs, as long as you are buying good brands and not random chinese crap.

While this somewhat generally true it's totally NOT in some cases. The difference between Samsung SSD 750 EVO 500GB and Samsung SSD 860 EVO 500GB and the WD Blue WDS500G2B0A is, in my opinion, not at all notable. The difference between those and the TOSHIBA Q300 240 GB SSD I have is huge as in massive and immediately noticeable, specially on writes. I absolutely do not recommend the Toshiba to anyone but your worst enemy. Perhaps
>buying good brands
covers this but I assumed Toshiba was a alright brand since their 3 TB HDDs are really good (fast and reliable but also a bit noisy).

[starshipeleven]

Perhaps
>buying good brands
covers this but I assumed Toshiba was a alright brand since their 3 TB HDDs are really good (fast and reliable but also a bit noisy).

Yeah ok, I should have not assumed everyone knew what "good brands" meant for me.

I don't regard Toshiba as a particularly good brand in the memory device space. I don't like them also for their laptops as they have some curious need to go out of their way to restrict access to service manuals https://www.techdirt.com/articles/20...computer.shtml (meanwhile others like say HP/Dell/Lenovo hosts them on their site, or like most others does not give a damn about people collecting them in a website), but we are getting sidetracked.

Toshiba is using OCZ know-how for SSDs (and OCZ was kind of crap and went bankrupt for a reason) after they bought them for peanuts, so just as I avoided OCZ I keep avoiding Toshiba. I also never liked their 2.5'' hard drives as they usually required twice as much power as a comparable WD or HGST.

Good brands imho (those I have positive experiences with) are Crucial, Sandisk and WD (which bough SanDisk for like 19 B$ and is rebranding their SSDs ever since).

Samsung is Ok for performance by they had too many SSD firmwares that were flaky and required features to be blacklisted in the kernel, and I also don't like them for other reasons not related to storage devices so I try to avoid them if possible.

[Markore]

I wonder why ZFS (ZfsOnLinux and/or Ubuntu's ZFS kernel module) isn't tested for the same benchmark?
Obviously all top VDEVs added to ZFS pool will do striping across all the pool. Comparing that much speedup with other Filesystems would be interesting, providing ZFS does serialization of operations. (Yet there is more job to do, not unilike Btrfs)