NVMe SSD raid-z1 shows low throughput, limited by compression or just bad benchmarking? #16930
Unanswered
LunarLambda
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Built a raid-z1 out of 3x Samsung 990 PRO 2 TB (2 GiB DRAM cahce) NVMe drives.
With fio doing random I/O on 4 threads I hit about 160 MB/s read and write, and in rudimentary
dd
tests I hit 2.4 GB/s for write and 1.2 GB/s for read.zfs settings are
relatime=on xattr=sa dnodesize=auto compression=zstd normalization=formD acltype=posixacl
. Otherwise everything's default forzfs 2.2.7
on Linux 6.6.69.ashift=9
as the SSDs only support 512 sectors. I knowashift=12
might be better but I don't imagine I'm losing /that/ much to it.The SSDs are capable of much more (5-7.8 GB/s advertised), and are in a decently new machine (AMD 9800X3D, 32 GiB 6400 MT/s RAM), so either
zstd
is really cutting my throughput down or my testing methodology just sucks.I'd love to know a more rigorous way to test what performance I can get out of my zpool, or even just anecdotes about what kind of performance you can expect from a typical ZFS setup relative to the drives' rated speeds.
EDIT: I also threw kdiskmark at it with the NVMe SSD preset and that gave significantly better looking numbers:

If this is more representative of actual throughput I can see this being down to use
zstd
.Beta Was this translation helpful? Give feedback.
All reactions