I'm going to use EXT4 (with type small because my custom distributed database writes many small JSON files) and hope for the best on my Atom 8/16-core machines that have 12 SATA ports.
I wont use raid but instead plan to mount these "manually".
Would love some feedback! :S
Good point! Unless we have a EMP! ;)
Edit: I found my M2 spacers I ordered by mistake 5 years ago, they are a bit wide at 18mm but that is good for thermals! Genious!
Edit2: Ordered 6mm spacers! You can get an idea how it will look here: http://move.rupy.se/file/SSD_unit.png
More important than cooling is the voltage rails. Reliable PSUs that don't overvolt the load under any scenario are important.
Will use Mean Well from 220V -> 12V... but after that it's PicoPSUs! (160/180/200W I will try many to see how their 3.3/5/12 look, I think this drive uses 5V but how can I find out? also measuring voltages on a live circuit has burned me before how do you make sure not to short/fry stuff?!)
Another thing I had to consider was power since I plan to back these up with lead-acid batteries for power outages. These drives draw 2W when active... so that is 18W for 8x (512GB) DB drives and one OS drive! Compared to the 25/32W of the 8/16-core motherboards that's alot... but mechanical drives are insane on electricity, up to 10W per drive!!!
But good point, do you know what connector tese use?
If you're concerned about battery life, you really shouldn't be using lots of small drives, especially when the controllers are so ancient. You'll get at least 5x the performance per Watt from a single newer flash-based SSD and much better idle power behavior. Consumer SSDs also offer much more in the way of low-power idle states than enterprise SSDs. At the extreme, you can get a $75 512GB drive that provides significantly more performance than 8x X25-Es in aggregate, while never reaching 3W power draw under load and idling at under 10mW.
To me data integrity and longevity are paramount!
These have same idle power btw, but ofcourse 2W on read/write is bad but I rather use these $1000 in 2011 50nm SLC industrial drives and compress my database than buy some MLC/TLC/QLC drive made today with 14nm!
Unless you have measured your own workload in enough detail to know otherwise, it is extremely unlikely you actually need to use SLC NAND. The market has almost completely abandoned SLC NAND for good reason, and your use case probably isn't that special (though if it is, I'm sure we could have an interesting conversation about it). Absent clear evidence to the contrary, I'm going to assume that a modern TLC SSD could easily fulfill your actual write endurance and data retention requirements, because that's the conclusion that has been reached by all of the big corporations who have been using modern SSDs in production at scale.
That said these were high transaction server workloads, and typical home use would be challenged to get to that kind of wear.
Some companies still make new "psuedo-SLC" drives [1] which store a single bit per flash cell. The downside is they're expensive: $1/gigabyte instead of $0.1/gigabyte.
And at that price, RAID mirroring starts to look like a great deal!
[1] https://uk.farnell.com/c/office-computer-networking-products...
These drives seem to be built like tanks and they fill the 2.5" enclosure to the brim.
They are also the first drive with RAM buffer on the read lane which is interesting in it's own right; meaning the motherboard never have to wait for writes but the disk have to wait for the motherboard on reads!!!