Disk performance problem zfs raid1 vs hardware raid1

[pisarz77]

Hi,
I test new servers, 2 identical machines Dell T550 with 2x1tb ssd with PERC H355.
On pve1 I set hardware raid1 and install proxmox with default lvm.
Pve2 I set both drives to non-raid in controller and have installed proxmox with zfs raid1.
On both proxmox i start identical windows vm (clone) and run CrystalDiskMark test. Here are results:

pve1 hw raid1:


pve2 zfs raid1:


Question:
Is there anything I can do with zfs to get closer to hwraid1 results ?
Some tweaks or changes in configs or whole concept?

I'm looking in zfs for replication but with such difference in write speed i think its better to stay with lvm and forget replication ?

Any ideas much appreciated.

 

[floh8]

This is the reality. ZFS is a cow-Filesystem with its benefits therefor it is slower.

 

[Dunuin]

This is the reality. ZFS is a cow-Filesystem with its benefits therefor it is slower.

Jup, you get additional features and better data integrity when using ZFS but this comes with the cost of a lot more overhead so worse performance.
You have to decide if you want a slower but safer storage or a faster but less secure one.

You could try to optimize it by changing the volblocksize, ashift and relatime.

PS:
CrystalDiskMark is a very bad benchmark utility as it will only do async writes. Actual performance might be way worse, especially if your workload also consists of sync writes.
Better and way more powerfull would be fio but it's not easy to use it or to get useful results.

More useful would be the latency, so a 4K random sync write test with depth of 1 which would show the worst case scenario which CDM can't do.

 

[LnxBil]

CrystalDiskMark is a very bad benchmark utility as it will only do async writes. Actual performance might be way worse, especially if your workload also consists of sync writes.
Better and way more powerfull would be fio but it's not easy to use it or to get useful results.

YES! 29 GB/s is very VERY WRONG for two SSDs. Even the newest generation of NVMe are NOT capable of performing this good.