
"Anyone fancy testing the 'unlimited' drive writes claim on Nimbus Data's 100TB whopper SSD?"
Yes, sure, send me a couple. You have my address.
Nimbus Data has introduced its 100TB ExaDrive DC series SSD, the highest-capacity flash drive available. It has some unusual characteristics but, first, let's show its size advantage by checking out the competition: Toshiba PM5: 30.72TB, 2.5-inch case, 12 Gbit/s SAS interface, 1, 3, 5 and 10 drive writes per day (DWPD) …
FUCK YES! Gimmiegimmiegimmie!
How many hours of "adult content" do you have if you want one of these?
By my reckoning that's about 50,000 hours of skin-flicks, and if you watched a full eight hours a day, it'd be the year 2035 before you'd finished watching.
Speed "indicates that online transaction processing-type applications are not for this drive."
Actually, once a given transaction is completed and its data set calms down and is ready for archiving (after about a minute), then the associated data can be moved to slower drives. You'd think that such an system array could be 99.9% slow drives and 0.1% ultra fast drives to ensure that the active transactions are kept on non-volatile storage and properly journaled at every step.
Hopefully they're not dealing with 100TB of transactions every few minutes.
"The same ExaDrive-sourced capacity [100PB] would need one rack, 990 drives and draw about 16kW"
These are 3.5" drives, according to the article. How exactly are they planning to fit 990 of them into a single rack?
Let's be very optimistic and say there is 48U usable space in the rack (no power distribution, no networking) so you can fit 12 x 4U chassis in there. They still need to fit 83 x 3.5" drives per chassis.
Even the Backblaze storage pod v6 only does 60, with disk utterly jammed in - and it overhangs the end of the rack by over 4 inches.
And the cost... well let's just not go there :-)
Agreed... I was also considering the power footprints. On a drive that takes about 40 hours to read completely, I was thinking that the drive has no purpose other than cold object storage. That being said, 16KW seems a bit silly. Why the hell would anyone consider keeping all these drives powered at all times? If they optimized the boot performance of the drive, it should be possible to leave these drives powered down except when needed. As the data gets older, it seems that data can stay offline for possibly months or years at a time. If there are 60 of these in a single 4U enclosure, most hyper scale companies couldn’t generate 4U of data a year. 6PB is actually quite a lot. Of course, these would be mirrored at least three ways in different locations.
But in the end, 16KW to power a rack... possibly for years at a time seems like a really bad idea. It’s not really a valuable measurement to compare a “wasting power” footprint.
A 7U module would allow for front-loading 51 drives (3x17) and a rack could hold 6 of them, which is 306 drives. However, other than making hot swap difficult there's nothing stopping you from using the full rack depth, which would allow 1836 drives in a standard 42U one meter deep rack.
OK, you probably need a few bays for electronics to connect to all those drives, but there's plenty of room to pack them a little less than maximally tight and still have room for backplane, controllers and fans, and reach 990 without maximal packing density. Not suggesting its a good idea, but its possible.
"These are 3.5" drives, according to the article. How exactly are they planning to fit 990 of them into a single rack?"
1: How deep is your rack?
2: How wide is your rack?
I have 1.2 metre deep Netshelters at the moment and I can't be the only one who remembers 1.6metre deep PR1ME cabinets.
19 inches isn't the only width you can get - 22 and 26" are also available.
Even at 4U 19" - see https://www.ixsystems.com/ix-server-family/server-jbod-enclosures/ and the ix4090jtl
Here's the thing about these ultra high-speed drives . . . what are you putting them in? If it's a standard controller-based array, the controller will run out of performance before your drives do, so an array of lower-speed drives will actually serve most workloads just fine. If you're using NVMeOF, that may be a different story, but for a lot of workloads, this drive will be perfectly adequate and massively superior to old-school spinning rust.
Bear in mind these things wont be sitting in some rack or some server somewhere waiting to be filled before moving onto the next one. They would be in an array where the controller will manage redundancy and writing to multiple SSDs at once, so your "write speed" for the array would exceed the 500MB/s.
This is assuming (which is unlikely) that all the SSDs in the array are all of this type and faster discs aren't sitting in front caching before it writes to the slower discs.
PS give me one (just one) to test. My internet speed at home wont be able to write at 500MB/s so no bottlenecks on the SSD.
Put the drive into a write/read loop over a 100mb area. Write that one area over and over an over again. Such would be a realistic actual use condition. at 500mb/s it would be 5 writes per second, 300/minute, 18,000 per hour, 432,000 per day, 3,024,000 writes in a week..... See how long the drive survives in this test.
Comparing a 100Pb rack an saying Sammy's offering would require 6x45U racks may be true but they're 2.5" units, how big is the equivalent Nimbus racking based on 3.5" units, perhaps half the size? Certainly not one third as inferred by simple "capacity per unit whilst ignoring the physical size" calculation ...
I would like to test, ad nauseum, the Nimbus Data's 100TB SSD; I think it would be necessary to test the write/read process in an optimised, sequential process (optimised against the operational characteristics of this drive) - and over the whole 'surface' of this drive.
It could then be said that - at the completion of each sequential cycle, 'n' number of write/reads had occurred on every write/read area of the storage (where 'n' is the number of sequential cycles)...
Keeping the drive going 24/7 for an agreed number of sequencial cycles to emulate the average number of power downs (after, say a number of trillion, quadrillion, quintillion write/reads or sequential cycles - could also give anothet realistic metric - until the drive fails (if it fails)...
etc., etc..
William