Hacker News Reader

Comment by bifrost | Hacker News Reader

SSD technology still has a reliability hurdle, you can nuke one in a month if you write to it constantly. Spinning media has much longer read/write durability. Until there is parity its still a bit of an arms race for capacity.

Hybrid drives are pretty awful so I don't think they'll stick around. I could see Hierarchical Storage Management make a comeback and an entry to the consumer space, but thats been esoteric at best.

View on HN

False.

http://techreport.com/review/27909/the-ssd-endurance-experim...

True, and I've done it! I accidentally used a consumer grade (Crucial M500) in an application where endurance would be reached in less than a month.

In fairness to Micron/Crucial, the drives did not lose any data.... but the write bandwidth degraded down to 10MB/s which counts as a failure for most apps. That's the catch in the report you cite: the bandwidth of a used-up drive is so poor that it's hard to actually write enough extra data to fully-fail.

Edit: Burning up an SSD was how I learned that TRIM is unnecessary: even though many parts of the filesystem were stable and never re-written, the bandwidth was even across the entire drive. The firmware noticed that some cells were underused and moved the stable data off them so it could use up the endurance evenly.

"In fairness to Micron/Crucial, the drives did not lose any data.... but the write bandwidth degraded down to 10MB/s which counts as a failure for most apps. That's the catch in the report you cite: the bandwidth of a used-up drive is so poor that it's hard to actually write enough extra data to fully-fail."

You call that a catch? I would sell it as a feature. It's way better than "looks OK until t minus 1, can't read or write from time t onwards" or lesser variants of the "try reading everything a few times, and you may get 90% of your data back"

I wonder how fast it would be to boot Windows with a SSD like this.

> The firmware noticed that some cells were underused and moved the stable data off them so it could use up the endurance evenly.

It moved data off them, but had to move it somewhere. I would have just rewritten them outright if it had known they were actually free.

Can you please elaborate on what application would basically kill an SSD in a month? It is my understanding that the Crucial M500 drives support something like 72TB of writes.

Please don't take any offense, I'm just interested in how you were using the drive.

Continuous record of 8 streams of 220Mb/s video. There were 10 SSDs for storage in a RAID-6 configuration.

So the effective endurance of the RAID was (10 - 2) * 72TB = 720TB.

8 Streams of 220Mb/s is 220/s, 13.2GB/minute, 792GB/hour, 19TB/day. So about 38 days.

While you are correct that you can nuke a normal SSD in a month, you can buy high-endurance drives which will last much longer. For example, you can buy a p400m with 7.5PB of write endurance [1]. It's just a question of making sure you buy a part that matches your need.

If you're writing more than 7.5PB in a month... I'm prepared to be amazed :-)

[1] http://www.tomshardware.com/reviews/p400m-ssd-enterprise,342...

Endurance is really not a concern for the vast majority of consumers. The media wearout indicator on my 19-month old cheap TLC drive has dropped 2%, and I'm a fairly heavy user.

The vast majority of SSD failures seem to be a result of firmware faults.

Do they? The article specifically says otherwise and links to this:

http://itknowledgeexchange.techtarget.com/storage-soup/netap...

It also mentions that the 3D NAND process dramatically improves matters over previous tech.

I'm not saying you're wrong, but it's weird that you don't acknowledge that the article addresses your point.

https://www.nimblestorage.com make what's essentially a flash+disk hybrid HSM-in-a-box. They're very nice, from experience.

Reply on news.ycombinator.com