sync = disabled vs SSD ZIL: too much difference
Posted: 18 Jul 2016 20:51
At home I have an All-in-one server based on Supermicro X10SL7-F MB, practically something very similar to this: https://b3n.org/freenas-9-3-on-vmware-esxi-6-0-guide/
I know that the performance of the combination ESXi + NFS + ZFS can be disastrous.
So to improve them I added to my ZFS pool (2x2TB SATA HDD in mirror) a Samsung SSD 840 EVO over-provisioned at 8GB (https://www.thomas-krenn.com/en/wiki/SS ... ing_hdparm) as SLOG/ZIL.
The result was very disappointed.
CristalDiskMark in Windows 2008 R2 x64 VM reports:
It seemed to have read that by doing so you could get "comparable" performance with sync = disabled, right?
Later I tried with a second SSD, the first mirrored, then striped, and also increased the number of NFS server threads, from 4 to 32 (https://b3n.org/freenas-vs-omnios-napp-it/) but without significant improvements.
So I again removed the SSD from the pool, I set sync = disabled and I redid the test:
Does it seem possible?
In the end I created a mirror pool with only the two SSDs and I was literally speechless:
The write performance are absurd!
I do not understand!!!
Note that all HDDs are connected to the internal controller LSI 2308, while the two SSDs are connected to two SATA3 ports of the integrated Intel Lynx Point controller, both passed through to N4F via VT-d.
I know that the SSD is not an Intel S3500 but it is possible that the difference is due to the kind of SSD?
Where can I compare the performance obtained from other, perhaps with configurations similar to mine?
Any other suggestions?
I know that the performance of the combination ESXi + NFS + ZFS can be disastrous.
So to improve them I added to my ZFS pool (2x2TB SATA HDD in mirror) a Samsung SSD 840 EVO over-provisioned at 8GB (https://www.thomas-krenn.com/en/wiki/SS ... ing_hdparm) as SLOG/ZIL.
The result was very disappointed.
CristalDiskMark in Windows 2008 R2 x64 VM reports:
It seemed to have read that by doing so you could get "comparable" performance with sync = disabled, right?
Later I tried with a second SSD, the first mirrored, then striped, and also increased the number of NFS server threads, from 4 to 32 (https://b3n.org/freenas-vs-omnios-napp-it/) but without significant improvements.
So I again removed the SSD from the pool, I set sync = disabled and I redid the test:
Does it seem possible?
In the end I created a mirror pool with only the two SSDs and I was literally speechless:
The write performance are absurd!
I do not understand!!!
Note that all HDDs are connected to the internal controller LSI 2308, while the two SSDs are connected to two SATA3 ports of the integrated Intel Lynx Point controller, both passed through to N4F via VT-d.
I know that the SSD is not an Intel S3500 but it is possible that the difference is due to the kind of SSD?
Where can I compare the performance obtained from other, perhaps with configurations similar to mine?
Any other suggestions?