I currently use 3 NVME drives (Samsung 980 Pro 2TB) as a zfs “stripe”. One hangs of the CPU PCIex4 lane, the other 2 off the chipset. So effectively one has 4 lanes, the other 2 share 4 lanes (along with any other odd peripherals on the chipset).
What is the real-world performance hit compared to all on 4 lanes each. I could buy a riser card and put the 2 on a PCIex8, but that does cost half a single NVME, so does not seem very cost effective…
Please share any practical experience - especially if you have tried the same upgrade! Thx
Any experience
An NVME stripe with ZFS is nonsense, unless you need the features of ZFS, except data integrity. If you’re primarily interested in performance, go with MDADM.
To your question, whether AMD or Intel, if you halve the interface bandwidth, given your NVME can actually use the full bandwidth, you will only get 50% of the performance. I mean, what did you expect?
@Janos - thanks for taking the time to reply. I am indeed interested in the features of ZFS (and all my systems from Raspberry Pi up are using ZFS), and I am happy in this case with the lack of data integrity.
A PCIe gen4 lane is about 1900 MBps and the SSD 7000 read/ 5100 write (best case). So i realise that there will be a hit for very large transfers. However, for day-to-day use with a variety of reads and writes, elements cached in RAM and all the system foibles (different latency due to CPU/Chipset, OS activity etc.) , I’m interested in the actual impact.
I think at the moment with Open ZFS 2.1 you won’t see any relevant difference, except for things you’ve already mentioned, but that might change soon if Direct-IO gets implemented with Open-ZFS 3.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.