It might also make sense to try setting NOCOW ['chattr +C'] on the files/directories used for database testing, as this would normally be a recommended production configuration due to the impact of fragmentation (although whether most SQLite databases are set like that is an interesting question).
Announcement
Collapse
No announcement yet.
Benchmarks Of Btrfs RAID On Four Samsung 970 EVO NVMe SSDs
Collapse
X
-
I am waiting for the comparison with ZFS !
Also, for single and dual drive, please include benchmarks of on-board m.2 slots to compare with the performance of the XPANDER-AERO. Just in case that there is performance degradation with the pcie board.
Also, do you have intel optane drives on hand ? That would also be nice to see !
Thanks
Comment
-
Michael could you modify the database tests on Btrfs to apply the C attribute on the test directory using chattr, before initiating the test? It's to disable the CoW system for that directory and to my understanding it's the best approach to handling these directories on a CoW filesystem. It would be interesting to see what the results would be.
Comment
-
Originally posted by schmidtbag View PostHaha I meanwhile am just interested in where I can get my hands on the XPANDER card without buying the motherboard. That thing intrigues me.
MSI seems to be a latecomer.
Comment
-
Originally posted by R41N3R View PostI do use btrfs a lot, but actually always with noadata and with chattr +C for databases and VM's. I think the test with defaults is quite unimportant, but at least there are some numbers to look at ;-)
However, it would be interesting to also adapt certain "tuneings" that many BTRFS users most certainly will apply such as disabling data-cow and isolating such parts of the filesystem (subvolumes). Once you got per. subvolume "raid" levels things really start to get interesting as well.
I also think that I should point out that some of the tests being done on BTRFS suffers from the fact that BTRFS choose what mirror to use based on the PID (process identifier). So let's say you run some sqlite read performance test from one process, for example with PID: 5432 then BTRFS will always pick mirror number 2 unless there is a read error. If you at the same time in parallell run a process with PID 4321 then you would read from both copies and get a significant performance increase.
Now , this is not (contrary to popular belief) much different than MDRAID, but you really have to understand how BTRFS work to interpret the benchmark results as well.
The good thing is that there has been patches posted that will make BTRFS pick a storage device based on the queue. The sad thing is that these patches has not been accepted (yet). Once those (or similar patches) gets merged the defaults would be even more interesting to watch
http://www.dirtcellar.net
Comment
Comment