Earlier quoted context omitted.
Compression in ZFS doesn't use RAM, dedup tables, the arc and l2arc mapping tables are all that matters with regards to memory usage with ZFS. l2arc is where a lot of newbies end up shooting themselves in the foot, they go toss in a mirrored pair of 512GB SSD's for their l2arc on a system with 32GB of RAM and wonder why performance got worse or it crashed, or they enabled dedup on 24TB of data with 64GB of ram and wo…
Ah crap. The advice I got was just about the opposite - I wanted an L2ARC device in my system for performance. Atom C2758, 32GB ECC RAM, 4x4TB in RAIDZ2 (losing that much storage was painful , and the $1200AUD upgrade path to 4x8TB even more so) with a 128GB SSD as L2ARC. I found out some time later that I probably wanted an SLOG device instead but I'm really too afraid to touch my config. FWIW it's pretty stable on…
RAIDZ is a performance killer right off the bat, I would switch to using mirrored vdev's instead if throughput is an issue for you. Parity calculation kills write speeds, especially on slower CPU's like a C2758, and RAIDZ doesn't give you any extra throughput on reads since there's only one usable copy of a stripe to read from. I have 2x4TB disks in my pool as a mirrored vdev (with an additional 2x3TB coming after I get my 128GB flash drive for XenServer to free up my second 3TB disk again) and I get reads over 500MB/sec using 10Gig-E to my XenServer host, and equally fast writes. Disks are fairly cheap, RAIDZ is not a good solution if you need performance, and if you are using four disks in RAIDZ-2 you would lose the same amount of storage using two mirrored vdev's anyway (and save the increased potential of a rebuild failing due to multiple disk failures that is increasingly common on higher capacity drives).
If writes are a bigger problem for you, then, yes, a SLOG device will help - but any random SSD is not going to do. If you are using a SLOG ZFS expects it will not corrupt in-flight data following a power loss, even a "high-end" SSD like a Samsung 850 Pro or a Crucial MX200 will lose data in the event of a power failure. You also don't want a SLOG that isn't mirrored, if your SLOG is corrupted you just lost your entire pool. In addition you need a SSD with proper power-failure protection like the Intel DC series. Also, large writes skip the SLOG (>64KB) entirely, so if you are bandwidth bound (either from network or to disk) it is not going to help, if you are IOPS bound it can help tremendously.
Also, you mention 4x4TB and 4x8TB, I assume your issue is the available disk bays in your system? I personally ignore high-capacity drives as they are far too expensive, and my HP ML10 that I run FreeNAS off only has 1 internal drive bay (with a $50 add-on I can buy to install an extra 3), instead of dealing with internal drives as well as limited capacity I bought a cheap DAS array (Lenovo SA120, here's a picture of mine and my cheap TP-Link layer 3 switch http://i.imgur.com/eEvtP6Z.jpg) that cost me $200USD and connected it with an equally cheap LSI 9200-8e SAS HBA ($40), I now have 12 hot-swap drive bays and can replace my FreeNAS box without worrying about how many bays it has, if I need more I just need to buy an additional enclosure and daisy chain it off the first. This isn't a dirt cheap solution by any means (my homelab gear is easily worth over $1500 at this point, including my Lenovo TD340 acting as my XenServer host), but it will save you money in the long run by allowing you to easily buy many cheaper drives than fewer more expensive (but high capacity) ones.
