zfs raid perfomance

aneubau

Renowned Member
Sep 5, 2008
44
1
73
Vienna/Austria
Hi,
I have here a huge performance difference in using SSDs for a zfs raid configuration:
We are using two identical server (Thomas Krenn Server-Tower Intel Single-CPU TI106S with
Supermicro X11SSH-F and 8x SATA-3 (6 Gb/s) SW-Raid Controller on Board C236).
System root is installed on zfs raid1 with 2x Samsung SSD PM863a 240GB.
VM data was on a zfs raid10 with 4*4TB Enterprise grade SATA HDD.
After running two years and seeing SSD prices really lower I decided to exchange the HDD for VM data by SSDs.
First I bought 5x 2TB Samsung SSD 860Pro for building a zfs raidz1 on one Server.
As prices still were dropping I had a change to get 4x4TB Samsung SSD PM883 for building a zfs RAID10.
Comparing these two systems show a really amazing difference between them, even when rearranging the Samsung SSD 860 Pro as zfs raid10:

System 1 (SSD PM863a, Samsung SSD PM883):
pveperf /rpool/ROOT/

CPU BOGOMIPS: 24000.00
REGEX/SECOND: 3155443
HD SIZE: 180.31 GB (rpool/ROOT)
FSYNCS/SECOND: 6691.77
DNS EXT: 83.49 ms
DNS INT: 1.42 ms

pveperf /tank/vmdata
CPU BOGOMIPS: 24000.00
REGEX/SECOND: 3112874
HD SIZE: 3556.48 GB (tank/vmdata)
FSYNCS/SECOND: 6525.11
DNS EXT: 87.03 ms
DNS INT: 1.49 ms

System 2 (SSD PM863a, Samsung SSD 860 Pro):

pveperf /rpool/ROOT/
CPU BOGOMIPS: 24000.00
REGEX/SECOND: 3402691
HD SIZE: 172.30 GB (rpool/ROOT)
FSYNCS/SECOND: 6979.83
DNS EXT: 100.43 ms
DNS INT: 1.86 ms

pveperf /tank/vmdata/ (zfs RAIDZ1)

CPU BOGOMIPS: 24000.00
REGEX/SECOND: 3315807
HD SIZE: 7379.57 GB (tank/vmdata)
FSYNCS/SECOND: 767.28
DNS EXT: 73.96 ms
DNS INT: 1.63 ms

pveperf /tank/vmdata/ (zfs RAID10)
CPU BOGOMIPS: 24000.00
REGEX/SECOND: 3416612
HD SIZE: 3689.00 GB (tank/vmdata)
FSYNCS/SECOND: 728.48
DNS EXT: 104.19 ms
DNS INT: 1.44 ms

Pools were created with:
zpool create -f -o ashift=9 tank raidz1 /dev/sdc /dev/sdd /dev/sde /dev/sdf /dev/sdg
zpool create -f -o ashift=9 tank mirror /dev/sdc /dev/sdd mirror /dev/sde /dev/sdf

(I did also test the system with parameter ashift=12 for the Samsung SSD 860 Pro, but no difference either.)

Are the performance difference between consumer grade SSDs (860 Pro) and enterprise SSDs (PM883) really that huge ?
Why is there no difference in performance between an zfs RAIDZ1 and zfs RAID10 ?
 
In general, please use a good tool like fio for performance measurements. Please read the still unfinished page:
https://pve.proxmox.com/wiki/Benchmarking_Storage

Are the performance difference between consumer grade SSDs (860 Pro) and enterprise SSDs (PM883) really that huge ?

Yes. Consumer grade are really not that good for a lot of IOPS/fsyncs. Overview here:
http://www.sebastien-han.fr/blog/20...-if-your-ssd-is-suitable-as-a-journal-device/

Why is there no difference in performance between an zfs RAIDZ1 and zfs RAID10 ?

There should be.