Improve FreeNAS NFS performance when used with Proxmox

TL;DR: zfs set sync=disabled your/proxmox/dataset

Lately I’ve been playing around with Proxmox installed on an Intel NUC (the cleverly named NUC6CAYH, to be precise), and I must say it is really, really, cool.

I usually store my containers and VMs on the local 180 GB SSD that used to be in my old MacBook Pro, since it’s reasonably fast and it works well, but I wanted to experiment with NFS-backed storage off my FreeNAS box (4x4TB WD Reds in RAIDZ1, 16 GBs of RAM, an i5–3330 processor).

Frankly, I was pretty unsatisfied with the performance I was getting. Everything felt pretty slow, especially compared to the internal SSD and, surprisingly, to storing the same data on a little WD MyCloud (yes, the one with the handy built-in backdoor).

My very unscientific test was creating a fresh container based on Ubuntu 16.04, and upgrading the stock packages that came with it. As of today, it meant installing around 95 MB’s worth of packages, and a fair bit of I/O to get everything installed.

The task was completed in around 1’30″ with the container on the internal SSD, 2’10″ on the WD MyCloud, and an embarassing 7’15″ on the FreeNAS box.

After a bit of googling, I came to an easy solution: set the sync property of the ZFS dataset used by Proxmox to disabled (it is set to standard by default).

The complete command is zfs set sync=disabled your/proxmox/dataset (run that on FreeNAS as root or using sudo).

To be honest, I don’t really know the data-integrity implications of this flag: both machines and the switch between them are protected from power failures by two UPSs, so that shouldn’t be much of an issue.

Anyway, just changing that little flag signlificantly reduced the time required to complete my “benchmark”, bringing it down to around 1’40″, very close to Proxmox’s internal SSD. Again, at the moment I don’t really need to run VMs/CTs off the FreeNAS storage, but it is good to know that it is possible to achieve much faster performance with this little tweak.