Can raid-z be hardware accelerated??

[kaede15]

Hello, first post here, I'm planning to move my current File server (areca 1231ML, win2k8, 10x wd500) to FreeNas. The main reason is for the zfs, cuz I'm planning on adding some new hdds of greater capacity but I don't want to be tied down with the conventional "same disc capacity" requirement.
My main question is can I take advantage of the Areca XOR engine to calculate zfs parities?? I guess is pretty waste if I just set it to JBOD..

thx

 

[survive]

Hi kaede15,

ZFS will happily use drives bigger than your existing 500GBs, but it will treat them as 500GB drives.

You can't use the Areca card to off-load any of the math for calculating parity, but that's really what's nice about ZFS! Modern processors are so darn fast these days that it really has obviated the need for "real" raid controllers with the dedicated hardware. My suggestion (if you are certain this is how you want to proceed) would be to hock the Areca and get a couple of LSI SAS controllers (you can find IBM branded LSI 2008 chipset cards for around $75 on ebay, look for the M1015 cards).

-Will

 

[kaede15]

Thx Will!!! Guess I'm confusing with the Beyond-Raid function from Drobo. I always thought that zfs's zpool can be made of different hdd capacities and can actually put to use some of those wasted spaces like Beyond Raid... Guess that is still Drobo's domain. Regardless I will still going to build a freenas box with zfs, just for the fun of it, and just as you said it's pretty cheap too...

I see your specs you have a raidz2 setup, and I don't know how much data you have but can you tell me how long does it take to rebuild in your case? Does it have performance hit said being raidz instead of raidz2? What is the read and write speed (real world) through a gigabit, jumbo framed enable network?

thx

 

[lrusak]

My read and write speed through a gigabit network is ~120MB/s for my Z2 array

 

[survive]

Hi kaede15,

You have to remember where ZFS came from....Sun wrote it for use in their "big iron" servers where you would add drives by the 10's or 100's, not one by one.

I honestly don't know how long it would take to rebuild from a failed drive on my setup...I've never had a drive fail in this rig. I know one of the nice things about ZFS is it only needs to rebuild the data that's on the drives from parity, so if the drives are half full it should take half as long as a traditional hardware raid card. Personally I don't do drives bigger than 1TB simply because the rebuild times get to long for comfort....

There is certainly some performance hit doing raidz2 over raidz, but I think it's pretty academic because processor power is so cheap...in my case my disks are far faster than my network connection so I ultimately lost performance I could never use in exchange for far better data security.

I don't use jumbo frames on my setup because I think they are pretty much a relic from back in the day when it took a whole processor just to drive the I/O from the NIC. Back then it made sense to increase the packet size to cut down on the number of interrupts generated by the system...today, with modern NICs and multi-core processors, not so much. I want to talk to other things besides the filer from my clients and the non-standard packet size can jack that up.

-Will

 

[xbmcg]

... you can put your zfs in a Corvette and have it hardware accelerated to 200mph.... ;-)

But yes, there are ways to accelerate things by hardware. Use more drives in parallel in a striped config (eg raid10) - this speeds up read / write, don't use parity - so no parity calc necessary, if you use raidz / raidz2 align disc count for data to a power of 2 + count of parity drives:

raidz1: 3, 5, 9, ...
raidz2: 4, 6, 10, ...
raidz3: 5, 7, 11, ...

it makes calculations easier and faster.

You can add dedicated fast cache, log drives (SSD's), log drive should be mirrored to prevent data loss in transit.
You can use more controllers in parallel, to offload the traffic from a single chip.
More memory = more cache = faster,

Add another NIC interface and bundle it using link aggregation to speedup the network. You can go for a faster NIC's & switches (10Gbps,...).

You also can leave a side all other resource-hungry stuff like deduplication, encryption, compression - so the system can concentrate on serving data.