No one wants over 16TB per drive SSDs

[sechs]

As it turns out, 16TB is already a lot of data to lose in a failure and takes a long time to rebuild in an array....
https://www.anandtech.com/show/14081/the-reality-of-ssd-capacity-noone-wants-over-16tb-per-drive

 

[LunarMist]

Koobface needs more than 640KB of RAM either.

 

[Stereodude]

As it turns out, 16TB is already a lot of data to lose in a failure and takes a long time to rebuild in an array....
https://www.anandtech.com/show/14081/the-reality-of-ssd-capacity-noone-wants-over-16tb-per-drive

Yeah, I don't believe that for a moment.

 

[LunarMist]

Surely the big datacentres like the Google or Amazon should be designed to assume a certain number of drive failures every day without reducing performance below an acceptable level or increasing risk beyond an acceptable level. Small businesses or individuals with small numbers of drives are probably too risk averse and would not be as impacted by watts per TB and facility footprints, etc.

 

[jtr1962]

I can say as an individual I probably wouldn't want anything much over 4TB. The problem is backing it up. Some SSDs suffer from a data retention problem if left sitting unpowered for long periods. HDDs in my opinion have squeezed so much data on a platter that they're inherently unreliable, period.

 

[sechs]

Yeah, I don't believe that for a moment.

The statement is hyperbolic, but the sentiment is there.

If you think about it, there's a level of data density where it's simply not currently worth the cost in money and time to deal the probability of failure.

 

[LunarMist]

That is probably about 100TB.

16 TB and larger will be normal once the drives have higher capacity per platter due to the MUMR and HAMR.
The recent trend of simply adding platters is not helping transfer rates or performance in general.

 

[Stereodude]

The statement is hyperbolic, but the sentiment is there.

If you think about it, there's a level of data density where it's simply not currently worth the cost in money and time to deal the probability of failure.

Everything else being equal, spreading your data across more devices increases the probability of failure. An AID-0 array of 16x1TB drives is almost certainly more prone to failure than one 16TB drive.

 

[LunarMist]

Everything else being equal, spreading your data across more devices increases the probability of failure. An AID-0 array of 16x1TB drives is almost certainly more prone to failure than one 16TB drive.

The drive doesn't know what data is on it. Failure depends on the power on hours and to some degree the number of bytes W/R.
Assuming a 16TB does not have more than 4x the failure rates of a 4TB drive, then 16TB would be better.

In any case 16TB will be considered small when the 32TB drives are available in a few years.

 

[jtr1962]

The statement is hyperbolic, but the sentiment is there.

If you think about it, there's a level of data density where it's simply not currently worth the cost in money and time to deal the probability of failure.

In the case of SSDs, unlike HDDs, it's not like the larger ones have higher data density. They just have more NAND chips. I have more confidence in SSDs since we've increased chip capacity by stacking cells, instead of using smaller cells. The primary problem with large SSDs is when they fail, you lose a lot of data.

I'm also sure that 16 TB number will be a moving target. Eventually, we'll have much faster interfaces to SSDs, so the rebuild time after losing 16TB might be no more than after losing 4TB now.

 

[jtr1962]

Everything else being equal, spreading your data across more devices increases the probability of failure. An AID-0 array of 16x1TB drives is almost certainly more prone to failure than one 16TB drive.

It's probably more complex than that. The failure of an SSD is some function of the failure rates of the controller and the individual NAND chips. Since drives in any given generation use similar density NAND chips, an individual NAND chip is not more likely to fail based on the size of the drive. However, larger drives use proportionally more NAND chips. If the controlling factor is the failure rate of the NAND chips, then a 16TB drive probably has about the same reliability as 16 1 TB drives. If controller failure is the controlling factor, then the larger drive is more reliable.

 

[LunarMist]

Apparently this article is about SSDs, not HDDs. :doh: It's amazing the difference after getting home and have a 27' display compared to the cell phone.

 

[Stereodude]

It's probably more complex than that. The failure of an SSD is some function of the failure rates of the controller and the individual NAND chips. Since drives in any given generation use similar density NAND chips, an individual NAND chip is not more likely to fail based on the size of the drive. However, larger drives use proportionally more NAND chips. If the controlling factor is the failure rate of the NAND chips, then a 16TB drive probably has about the same reliability as 16 1 TB drives. If controller failure is the controlling factor, then the larger drive is more reliable.

There are more PCBs, more solder joints, more of pretty much everything to fail (except perhaps for NAND chips).

 

[ddrueding]

The biggest thing for me is the ratio of capacity and throughput. Can I fill/backup a drive in under 24 hours? The system is designed to handle a failure, even two, but I don't want to be left with my pants down for more than a day while it rebuilds.

 

[LunarMist]

A 16TB nVMe SSD would be resilvered in less than three hours. It's far faster than any hard drive. UBER is two orders of magnitude better than a hard drive and MTBF is better as well.

 

[sedrosken]

Personally my 8TB is currently plenty, but there might come a day when it's not. I already am something of a digital hoarder anyway -- I find it easier to justify than being an actual hoarder. I can't say I have a ton of experience with RAID setups, or with dealing with drive failures (though today Samsur's drive finally bit the dust, kind of surprising considering how low-hours of a unit it was, but there you go I suppose) but I given the language in use here I assume rebuilding the array does take an awful lot of time, so maybe they're onto something for the moment. Is nVMe RAID a thing yet? I imagine that would make the sting of failure much less of a problem.