How fast do modern SSDs die: Upcoming test

[Patrick]

I just ordered two Samsung 750 EVO 120GB

drives. They have a whopping 35TBW rated endurance. These drives are going to be sacrificed.

Here is the plan:

• Two drives, same system
• One drive filled to X% (simulate having an OS plus applications installed)
• Write to the drives and monitor status
• See if there is an appreciable difference between an empty drive and a partially full drive
• If we see both drives die at basically the same number of writes, then there is no difference, if they die at significantly different write levels, then we will have an answer.

Speaking to folks both at drive OEMs and system vendors pretty much everyone expects the full drive to experience issues before the empty drive. Alas, looking at folks that have done endurance testing, they seem to be using full drives so there are a lot of folks that are using best case scenario data to form their understanding.

If there are any folks who manage large numbers of client PCs, I am looking for a solid data point in terms of average fill rate/ amount. Is the appropriate fill rate 100GB of 120GB, 80GB of 120GB or something else?

Feedback appreciated.

 

[T_Minus]

Looking forward to the results

 

[DavidRa]

I have a customer with a 1K endpoint Windows 8.1 +Apps COE that (IIRC) baselines at 45GiB but with one of their apps balloons to 200GiB (and that app is on ... 40% of machines). My employer's is closer to 90GiB. My recently installed VDI with Office, Visual Studio etc was 100GiB after install and Anniversary update; it dropped to 77GiB after a forced cleanup, and my everyday desktop is just over 110GiB.

The takeaway is that there's not really a good number that fits all users

 

[Patrick]

I have a customer with a 1K endpoint Windows 8.1 +Apps COE that (IIRC) baselines at 45GiB but with one of their apps balloons to 200GiB (and that app is on ... 40% of machines). My employer's is closer to 90GiB. My recently installed VDI with Office, Visual Studio etc was 100GiB after install and Anniversary update; it dropped to 77GiB after a forced cleanup, and my everyday desktop is just over 110GiB.

The takeaway is that there's not really a good number that fits all users

Yea I was thinking of using 95GB or 100GB static data as a baseline. The Samsung 750 is likely going to be sub 120GB formatted. Seems like that is not out-of-line. Even a user with a 45GB install may have local files.

 

[Cole]

Our measly ~150 machines average ~60GB usage. Mostly office machines. MS Office, Crystal Reports, etc. Most documents are stored on NAS.

 

[sean]

Assuming there is a difference in when they fail, it would be nice if you had drives at empty, 25%, 50%, 75%, and ~90% "static" data to compare the effects.

With "just" one drive, I think around 75% would be a good level. That would be a lot of data but not alarmingly so. That comes out to 90 GB.

 

[Patrick]

@sean sadly, limited funds for short term assets.

 

[GuybrushThreepwood]

Read the headline without reading "do". Went panic and shouted, what how??!

 

[Patrick]

More on this one tomorrow.

 

[wildpig1234]

why is that drive only rated at 35TBW? i thought these cells has 10000 rewrite time limitation so that should be 1.2PBW (120 * 10000)? My ocz vertex 3 120gb which is used as a boot drive on my win7 system has 23 TB written to them over the past 3 yrs according to ssdlife.

i am very interested in the outcome too... Although the way i use my array of drives is more like a video server/backup, in that i would write to a single drive filling it to capacity. when that drive is full, the next write is then to another drive filling that one up and then to another one and so on down the line in an array of 9 non raid HDD. When the last disk in that array is filled up, it's back to the first disk again.

I have a boot up SSD that's exclusive for that purpose with as much write limitation i can do to it as i can. the only thing on there is 20gb of os and apps. All the downloads and file saving are to the array of 9 HDD above. As i said, even though i don't intentionally save anything to the boot ssd, it still surprisingly got over 23 TB written to it over the last 3 yrs.

Thanks to a lot of background o/s transaction as well as our recently discovered web browser writes. it seems that a lot of time when i try to save file from a web page, IE would first save it to a temp file on the boot drive and then copy the file over to where i want it saved rather than download directly to a temp file which is saved right away on the chosen download destination .

So i guess in my scenario, the boot drive is the one that have the most stressed usage pattern?

 

[Patrick]

@wildpig1234 smaller process 2D planar TLC NAND. Finishing up a few photos on the piece.

 

[wildpig1234]

@wildpig1234 smaller process 2D planar TLC NAND. Finishing up a few photos on the piece.

Still, 35TBW sounds really low or conservative. that's really not much at all.

 

[Patrick]

Just published the test setup piece
Samsung 750 EVO's Ready for Sacrifice: A modern SSD endurance experiment

Project Kenko 01 is underway.

 

[keybored]

Does the Anvil suite of apps give you ability to capture SMART data changes over time? Or do you have any other tools that can do the same? You'll probably see signs of trouble earlier if you monitor SMART attribute value trends.

 

[wildpig1234]

so the warranty of 35TBW for 3yrs means 31GB/day but they only cover you at that for 3 yrs.. but maybe the rewrite endurance is much higher? b/c it doesn't make any sense otherwise.

if the rewrite endurance per cell fails after say just 4 yrs at 31gb/day that would be 44.8TBW over 4 yrs. spread that over 120gb means the overall rewrite number is just 374 times. so if you have a drive that only have 20gb free with only the free space being rewrite, then that means failure after 7.5TBW

But if you do constant write and delete test, with write speed of 200mb/s (low conservative speed estimate), that's 720GB/hr, 17TB/day nonstop. so failure in less than a day? ;(. did i screw up my math somewhere? that doesn't sound right....

 

[keybored]

...
if the rewrite endurance per cell fails after say just 4 yrs at 31gb/day that would be 44.8TBW over 4 yrs. spread that over 120gb means the overall rewrite number is just 374 times. so if you have a drive that only have 20gb free with only the free space being rewrite, then that means failure after 7.5TBW...

It all depends on how well wear-leveling is implemented, if at all. If drive's controller is able to cycle those blocks of NAND that have static data in and out of service so that they're not stuck with the same data all the time then you'll see higher longevity. If NAND cells with static data never get touched then you'll keep hammering on the same set of cells that are unoccupied and the drive will die with many of its NAND cells still in a good shape.
Many will remember the infamous Samsung 840 Evo which had horrendous performance issues with data that was written once and never touched. I suspect if they had good wear leveling implemented in those drives then this problem, which they evidently traced down to voltage decay over time in unused cells, would not have been that bad in the first place...

 

[Patrick]

Here is a super interesting data point since I am watching this run for the first bit.

At only 150GB written the write rates on the two drives:
Project Kenko 01 - 0GB Pre-Fill - 70MB/s
Project Kenko 01 - 90GB Pre-Fill - 27MB/s

 

[keybored]

Is this all sequential write I/O? 4K? or mixed?

 

[wildpig1234]

Is this all sequential write I/O? 4K? or mixed?

Must be mixed or small cluster write. that's way too slow for large file write. then again I am not surprised at the much slower speed on the fuller drive. firmware is probably trying to recycle and moving around the already occupied sectors to other sectors to try to even out the recycling.

 

[T_Minus]

My guess is that if those speeds are after running non-stop 150 GBW then they could make sense for seq. and too fast for random on a consumer drive. The consumer drives (most) aren't configured for sustained work loads, let alone write-heavy so they can never "catch up" with performance until there's downtime for GC and TRIM and whatever else software they have on them that makes them perform 'good' again.

I would hope after some downtime they would be back up to where you'd expect them, but I guess depending on the firmware may require a secure erase / full write format.