Torpedo at the Dock
The super fast disk project: Part-2
Wow, I just launched out into the SAS-RAID waters, and “BAM!”, I've already taken a torpedo within sight of the dock!
I ordered the 8-lane SAS-RAID controller (open box) from Other World Computing. The was the official start of the project. The big leverage of this solution was to be a really low cost-per-gigabyte in the RAID-0 array by using SSHD (solid-state-hard drives). Since I already have two of them, the plan was to re-purpose them and buy two more.
While the best-case cost of the fancy schmancy Samsung SM951 M.2 SSD works out to $0.68/GB for the 512GB model (smaller capacity raises the cost past $1.00/GB), I was hoping to get things down to $0.15/GB incremental capacity cost, and an overall cost (including the controller) of less than $0.34/GB.
The linchpin of the SSHD plan is that the 8GB cache in the SSHD's are effectively additive in a RAID-0 array, and that the cache hit rate would rise very quickly because of the relatively small working set of a typical project. The other half of that idea is the assumption that the data transfer rate of the cache would be approximately the same as a regular SSD. That's where it all went wrong...like, "BANG...gurgle, gurgle."
I have my first virus/cold in almost 3 years right now. While I was up at night hacking and sneezing, I decided to find out the actual transfer rate of the cache embedded in the Seagate 500GB drive. After reading literally hundreds of web pages of pablum, I found the gorgeously written analysis at techreport.com. In two particular entries, the small file reading test (read burst speed) that's intended to exercise the DRAM/cache of the drives showed that it has little more than about 1/4 the speed of a common SSD. That's right, in a RAID-0 array, the best I could hope for from these SSHD hybrid drives would be to match a regular single SSD. That, mon ami, is a disaster for the SSHD hybrid plan.
I blame the accountants! This problem has the distinct aroma of a great technological hot rod racer of an idea on which someone in accounting opted to mount square wheels; probably because square wheels waste less space in the warehouse. Every computer enthusiast that sees these drives immediately notes the 8GB cache as being too small. Most never find out that it's also too slow. There's an enterprise version of this product made by Seagate that includes 32GB cache on a 15K RPM 600GB drive. I'd buy 4 of them if they were $100 each...but they're not. For the budget-minded enterprise, there's a 300GB and 450GB version that undercuts the $1000 price of the 600GB version ($1.67/GB).
I'm willing to bet that the Seagate ST600MX0004 SSHD cache runs at full SATA-3 speed. HGST, Toshiba and IBM/Seagate have similar products. You can read all about it at the next link below. Viewer discretion is advised for the technologically queasy.
So, what now? The best way forward is to pick a target cost per GB and find a combination of SSD's that will match that target. With a 1TB (or 960GB) target, a combined cost of $0.30/GB that includes the $189 controller leaves only $101 to buy the SSD array. Nothing in the used market last night has 240GB SSD's at $25 apiece; normally they're $75 each these days. That puts the incremental cost at $0.31/GB (good) but the overall cost at $0.51/GB. That's not cheap.
Even ramping up to a 2TB (or 1920GB) target, the 512GB SSD's all cost more than $150. Even allowing for the fact that I'd only have to buy 3 of them that puts the 2TB total at $0.47/GB. This is not much better than the $0.51/GB of 1TB; most disturbing.
We're taking on water, but we're not sunk yet. It's time to restructure the paradigm of the plan.
- Ted Gary of TedLand
Sept 18, 2015