Unraid storage performance

porina · May 28, 2016

Tried reading around but didn't get far... I've got a HP microserver gen8 with 4x 3TB WD reds, set up with one parity disk and the rest for storage. No cache drive. CPU is Celeron G1610T and ram is 2x4GB ECC.

I've got a single windows share and am using that for backups of my main system. Like I saw when I tried the demo, there is a short time of near gigabit speed writes, then it'll drop to around 200-250mbps. I note the CPU usage is up to around 50% suggesting it is using a single core. I further note the disk activity counters increase roughly equally for read and writes on both the storage drive in use and the parity disk.

That leads me to assume, when you write, it reads both the old data on the drive and parity drive, mix with new data, to create the new values to write. So for every 1MB of new data written, it has to read 1MB from each of two disks before writing 1MB to those same disks. So my thinking is, are the disks limiting, or is it the CPU's ability to calculate parity?

It is not impossible to get a faster CPU but they're not cheap, and I'd want to be really certain before I go that route. Alternatively, if it is the disks, my only option then would be to use a cache SSD but I'm unsure on behaviour if the cache should fill before routine emptying. My backup strategy would be infrequent massive dumps, so worst case as far as cache is concerned.

leadeater · May 29, 2016

It's more than likely the CPU calculating parity that is slowing it down and SAMBA not using both cores. I would test this buy creating a mirror volume or one with no redundancy at all and do a copy test.

SAMBA which is used on Unix systems to create Windows Shares (SMB) isn't exactly that good/efficient which is why you are seeing all the load on the single core. If you create an NFS share you'll see that is does a much better job at load distribution and get better speeds.

porina · May 29, 2016

5 hours ago, leadeater said:

It's more than likely the CPU calculating parity that is slowing it down and SAMBA not using both cores. I would test this buy creating a mirror volume or one with no redundancy at all and do a copy test.

I'm pretty much all Windows so using NFS is not something I have a clue about. As for creating a mirror... I can't do it without breaking my existing array so that's not an option either. Just going to have to think about the cache drive option I think.

I'm in the process of doing my first time "everything" backup, and future incremental backups should be a lot more manageable! Worst case, the speed isn't a deal breaker since it is for backups only, but I still would have hoped for more.