Google's 6y study into enterprise and consumer grade SSDs - it's not pretty

[MageTank]

4 minutes ago, patrickjp93 said:

Except that the ratings have been found to be bogus, so the actual quality between the two is the same in this respect. Also, just because you read 1*10^15 bits from a drive that has a 1*10^-15 chance per bit to have an error does not mean you will wind up with an error. Do a little basic research on combinatorics and its role in probability.

I am opening a can of worms for even asking this... but can I have a source on the ratings being found as bogus? All I could find from google was a random reddit page (not exactly a trustworthy source for me). Aside from that, you are basically agreeing with what I have said, lol. "Having a chance for error". It's all luck. Having a higher rating means higher luck. Unless you are right about it being bogus, in which case, I'll need to see it to understand it.

[patrickjp93]

13 minutes ago, MageTank said:

I am opening a can of worms for even asking this... but can I have a source on the ratings being found as bogus? All I could find from google was a random reddit page (not exactly a trustworthy source for me). Aside from that, you are basically agreeing with what I have said, lol. "Having a chance for error". It's all luck. Having a higher rating means higher luck. Unless you are right about it being bogus, in which case, I'll need to see it to understand it.

It's in Google's report, a whole slew of testing.

[MageTank]

Just now, patrickjp93 said:

It's in Google's report, a whole slew of testing.

In the one posted in this thread?

[MageTank]

40 minutes ago, patrickjp93 said:

It's in Google's report, a whole slew of testing.

Still can't find that report. However, the original consensus has not been changed on Wikipedia yet ( I know, Wikipedia is not considered a valid source to most people, but they do provide sources for their claims in the bottom)

 

https://en.wikipedia.org/wiki/RAID#URE

 

It basically confirms exactly what I said before. Can you provide me with a link to your claims or not?

[Doobeedoo]

In earlier SSDs when they were still not mature tech, my first one, and that was the only that failed. So as far as reliability I'd say after years of improvements on SSDs it's all great, have couple of them and no issues. As far as endurance, endurance ratings don't satisfy me to much. I mean I won't exceed it within a month or so but still a lot of data is going through it and would like to see significantly more endurance in next gen SSDs.

Would be great if it could be like RAM not needing to worry how much data passed through and lowered it's lifespan.

[79wjd]

3 hours ago, Doobeedoo said:

In earlier SSDs when they were still not mature tech, my first one, and that was the only that failed. So as far as reliability I'd say after years of improvements on SSDs it's all great, have couple of them and no issues. As far as endurance, endurance ratings don't satisfy me to much. I mean I won't exceed it within a month or so but still a lot of data is going through it and would like to see significantly more endurance in next gen SSDs.

Would be great if it could be like RAM not needing to worry how much data passed through and lowered it's lifespan.

That raises an interesting question that I e never thought about before....why is the endurance of RAM so much higher than SSDs? (I assume if you wrote to ram enough you would eventually degrade the memory chips)

[2FA]

4 hours ago, MageTank said:

Still can't find that report. However, the original consensus has not been changed on Wikipedia yet ( I know, Wikipedia is not considered a valid source to most people, but they do provide sources for their claims in the bottom)

 

https://en.wikipedia.org/wiki/RAID#URE

 

It basically confirms exactly what I said before. Can you provide me with a link to your claims or not?

Are you expecting patrick to provide a source? I've personally only seen that happen once.

[MageTank]

Just now, DeadEyePsycho said:

Are you expecting patrick o provide a source? I've personally only seen that happen once.

I've seen it twice. Once from someone else, and once when provoked by me. However, I knew the chance of getting it this time was pretty slim. Sadly, I was kind of hoping he had a source to prove me wrong on this. Would be nice to know before investing in 8TB enterprise drives, seeing as I can get 8TB consumer Seagate drives for literally half the price.

 

Maybe the source for this google consumer vs enterprise HDD study is in Jim Keller's desk at the Intel headquarters that they are saving for him, and only he has the key to open it? That's my current theory as to why I have not been presented with any information after being so boldly called out (yet again). 

[leadeater]

1 hour ago, djdwosk97 said:

That raises an interesting question that I e never thought about before....why is the endurance of RAM so much higher than SSDs? (I assume if you wrote to ram enough you would eventually degrade the memory chips)

Yes this is also true for RAM but the write endurance on these types of memory chips is so high it's not even worth thinking about. Degradation due to age is a much bigger concern.

[79wjd]

12 minutes ago, leadeater said:

Yes this is also true for RAM but the write endurance on these types of memory chips is so high it's not even worth thinking about. Degradation due to age is a much bigger concern.

I figured that to be the case, but my question was more why is that the case. How does it work that the endurance is so much higher? Is it entirely different flash, or is highly binned, etc...? 

[patrickjp93]

5 hours ago, MageTank said:

Still can't find that report. However, the original consensus has not been changed on Wikipedia yet ( I know, Wikipedia is not considered a valid source to most people, but they do provide sources for their claims in the bottom)

 

https://en.wikipedia.org/wiki/RAID#URE

 

It basically confirms exactly what I said before. Can you provide me with a link to your claims or not?

Yes, in Google's report in this thread, which you already have the link for. Sorry for not responding sooner. Saturday is my workout day.

[patrickjp93]

1 hour ago, djdwosk97 said:

That raises an interesting question that I e never thought about before....why is the endurance of RAM so much higher than SSDs? (I assume if you wrote to ram enough you would eventually degrade the memory chips)

 

14 minutes ago, leadeater said:

Yes this is also true for RAM but the write endurance on these types of memory chips is so high it's not even worth thinking about. Degradation due to age is a much bigger concern.

No, no, no, no, and no.
 

DRAM is based on capacitors all lined up in a row, hence why it's volatile. They can't hold a charge more than a fraction of a second. NAND flash is an actual transistor-based design which can trap a charge state long-term. RAM can be written to and read from practically forever barring overcharge beyond the rated capacitance of any given bit-holding capacitor and barring regular entropy and the heat death of the universe. It just requires constant power.

[patrickjp93]

39 minutes ago, DeadEyePsycho said:

Are you expecting patrick to provide a source? I've personally only seen that happen once.

You haven't been around for a couple months have you?

[2FA]

Just now, patrickjp93 said:

You haven't been around for a couple months have you?

No I have been, I just don't follow threads to the point where the arguments break out. 

[leadeater]

5 minutes ago, djdwosk97 said:

I figured that to be the case, but my question was more why is that the case. How does it work that the endurance is so much higher? Is it entirely different flash, or is highly binned, etc...? 

Completely different. The way the data is stored is different since NAND is nonvolatile and RAM is volatile. Every cycle RAM is rewritten every time since data is stored in capacitors which slowly discharge.

 

http://computer.howstuffworks.com/ram.htm

[MageTank]

4 minutes ago, patrickjp93 said:

Yes, in Google's report in this thread, which you already have the link for. Sorry for not responding sooner. Saturday is my workout day.

I've skimmed through most of the 380 pages, and did not see any mention of URE's and their impact on large RAID arrays. Can you point me to any specific pages, or a specific word I should search within the document?

[leadeater]

33 minutes ago, patrickjp93 said:

 

No, no, no, no, and no.
 

DRAM is based on capacitors all lined up in a row, hence why it's volatile. They can't hold a charge more than a fraction of a second. NAND flash is an actual transistor-based design which can trap a charge state long-term. RAM can be written to and read from practically forever barring overcharge beyond the rated capacitance of any given bit-holding capacitor and barring regular entropy and the heat death of the universe. It just requires constant power.

 

Except RAM does fail, much more often than what is inferred by what you have written. Really can be bother debating this further but if you have worked for a company with multiple data cetners you will get a feel for how often ram actually does fail, it's more than practically never. 

[patrickjp93]

2 minutes ago, leadeater said:

 

Except RAM does fail, much more often than what is inferred by what you have written. Really can be bother debating this further but if you have worked for a company with multiple data cetners you will get a feel for how often ram actually does fail, it's more than practically never. 

TBs of RAM running in datacenters and even Petabytes in consumer space go on without ever failing before they're just thrown out with an old rig.

 

I interned for Google for a whole summer. 1 stick died out of 300TB worth, and that was because the board it was attached to died.

[79wjd]

2 hours ago, patrickjp93 said:

 

No, no, no, no, and no.
 

DRAM is based on capacitors all lined up in a row, hence why it's volatile. They can't hold a charge more than a fraction of a second. NAND flash is an actual transistor-based design which can trap a charge state long-term. RAM can be written to and read from practically forever barring overcharge beyond the rated capacitance of any given bit-holding capacitor and barring regular entropy and the heat death of the universe. It just requires constant power.

 

2 hours ago, leadeater said:

Completely different. The way the data is stored is different since NAND is nonvolatile and RAM is volatile. Every cycle RAM is rewritten every time since data is stored in capacitors which slowly discharge.

 

http://computer.howstuffworks.com/ram.htm

I understand that much, but I don't really understand why that makes a difference in the amount of times data can be written. Maybe there just isn't really an answer that would make sense to me as someone who doesn't understand how they work on an engineering level.

[patrickjp93]

6 minutes ago, djdwosk97 said:

 

I understand that much, but I don't really understand why that makes a difference in the amount of times data can be written. Maybe there just isn't really an answer that would make sense to me as someone who doesn't understand how they work on an engineering level.

It's very simple. For NAND flash, the voltage required to get it to change electrical states is higher than the silicon can "safely" allow passage. This is what wears flash cells out. You can charge and discharge a capacitor ad infinitum with no consequence as long as you don't push it out of spec. Think of NAND as more rigid and DRAM as much more flexible and thus the forces needed to change both being very different and effectual.

[RexinOridle]

Interesting article.

[TrigrH]

This study only has early ssds that are 4-6 years old, I doubt these issues are a problem at all for modern SSDs.

[ionbasa]

On 2/27/2016 at 9:29 AM, MageTank said:

I am opening a can of worms for even asking this... but can I have a source on the ratings being found as bogus? All I could find from google was a random reddit page (not exactly a trustworthy source for me). Aside from that, you are basically agreeing with what I have said, lol. "Having a chance for error". It's all luck. Having a higher rating means higher luck. Unless you are right about it being bogus, in which case, I'll need to see it to understand it.

Document page 67 of: https://www.usenix.org/sites/default/files/fast16_full_proceedings.pdf

 

Its right in the abstract:

Quote

The widely used metric UBER (uncorrectable bit error rate) is not a meaningful metric, since we see no correlation between the number of reads and the number of uncorrectable errors.

EDIT: That sections also goes on to explain the correlation and why the UBER is useless. See section 5.1 of that document. 

 

[allindaze]

I pray for my datas.

[Agent181]

this is why I do a double backup on hard drives and only keep applications on ssds since you can always install an app on an ssd but its easier to recover data on a hard drive, simple best of both worlds, oh and dont cheap out on an ssd, you get what you pay for.