Page 2 of 2
Re: NFS speed on 10Gb/s RJ45
Posted: 2018/07/24 10:30:33
by hunter86_bg
Did you notice any cpu or I/O bottlenecks when increasing the threads?
Edit: You may create a tmpfs(depends on your average file size) and export that and then try to read/write from the client. If you don't see any issues then it is I/O related.
Re: NFS speed on 10Gb/s RJ45
Posted: 2018/08/01 16:22:29
by Moe
Thanks everyone for the tips
I tried testing with a different tool ( fio instead of dd ) , and getting much better results
The commands I'm using are like :
Code: Select all
fio --name=read-test1 --ioengine=libaio --bs=9k --rw=read --direct=1 --iodepth=32 --runtime=60 --size=40G --group_reporting --filename=/mnt/volume/fio-read-t1
And I played with the blocksize from 9K all the way to 2M , 128k seems to give reasonable results
Re: NFS speed on 10Gb/s RJ45
Posted: 2018/08/01 20:38:37
by hunter86_bg
Poor performance with small block size indicates lattency issues, while poor performance with larger block size indicates bandwidth issues (or memory if bs is too large).
If you use 'dd' you need to test with 'oflag=direct' to get the 'truth'.
Have you tried to export tmpfs-based share?
Re: NFS speed on 10Gb/s RJ45
Posted: 2018/08/02 11:24:15
by Moe
hunter86_bg wrote: ↑2018/08/01 20:38:37
If you use 'dd' you need to test with 'oflag=direct' to get the 'truth'.
Yeah, I was using the conv=fdatasync
Have you tried to export tmpfs-based share?
No, I don't have access to the second server I was using as a client anymore, however I'm quite satisfied with what I got from NFS so far
Now I need to work on tuning a samba share