How bad will next gen NAND P/E cycle count will be?

[DirkGently1]

I have been following that thread over at XT, and it is quite remarkable how ALL the SSDs in that thread are coping so well with the punishment they are taking.

The endurance is even more amazing if the data in this graph is taken into account:

With the load these drives are under the nand is getting very little in the way of recovery time. Who knows how much longer it would last under normal desktop operation? Either way, we are way beyond having to worry about nand death any more.

If only the information were getting out to the people. I still see threads every day asking how to minimise writes. All because of the initial FUD in the very early days. Should the pagefile be moved..browser cache..etc, etc.

 

[=Wendy=]

The endurance is even more amazing if the data in this graph is taken into account:

With the load these drives are under the nand is getting very little in the way of recovery time. Who knows how much longer it would last under normal desktop operation? Either way, we are way beyond having to worry about nand death any more.

If only the information were getting out to the people. I still see threads every day asking how to minimise writes. All because of the initial FUD in the very early days. Should the pagefile be moved..browser cache..etc, etc.

I agree 100%.
SSDs (for most people) are still very new, and quite expensive. I guess people worry about wearing them out. The XT thread (IMO) shows very clearly that normal users should not worry about burning out the NAND on their SSD. Regarding moving files away from the SSD (pagefile). I agree that there is no need to do this, or any other file for that matter.

Folks should just enjoy their purchase and use them to the full.

The thread over at XT also dispels the myth that any other NAND other than the stuff that Intel and Micron uses on their (own) SSDs is crap, as the thread clearly shows the so called "2nd tier NAND" is every bit as durable as the stuff Intel and Micron are using on their own SSDs.

 

[hurleydood]

Go right ahead. I don't mind at all.



This shows the vertex write amplification doubled without TRIM.
You raised good point about things that make it vary but missed one crucial and obvious one... Data workload. I would imagine that if a drive is 90% full and is being given lots of tiny random writes (ex: log updates; a few bites in size they require writing a whole sector, so either 512b or 4kb depending on how it is formatted) then it might get to that theoretical 128x+ write amplification.
Suddenly 100 years expected lifespan becomes 1 year.

128x+ write amplification is real. Life span of 1 year is real.

My numbers indicate I have a write amplification of 175x.

3.2TB of lifetime data written with average P/E burn out of 4498 x 128GB.
562TB / 3.2TB = 175x write amplification

The only thing I do is browse the web and leave my computer on 24/7. Windows system logs, file system journaling, page file writes and browser cache are constantly writing small random data to the drive. I have plenty of ram and free space on the drive. All the small writes will kill your drive. I have replaced the drive recently as Windows finally started crashing even after I reinstalled. TRIM, garbage collection, and compression won't save SSDs.

 

[VirtualLarry]

Original Torx line has a 10-year warranty. Does that cover NAND wear-out? If it does, you should be set for a replacement.

 

[SSBrain]

128x+ write amplification is real. Life span of 1 year is real.

My numbers indicate I have a write amplification of 175x.

3.2TB of lifetime data written with average P/E burn out of 4498 x 128GB.
562TB / 3.2TB = 175x write amplification

Your Patriot Torx SSD has an old gen Indilinx controller, which is somehow able to get write amplification to horrendously high levels. I've seen several older Indilinx-based OCZ SSDs with similarly high WA values. Wear leveling on such drives must not be very sophisticated.

Modern SSD controllers with light usage typically have a much lower write amplification than that. On good drives 1.5-2.5x is the norm nowadays, not 175x.

 

[hurleydood]

There's absolutely no garbage collection. My bet is all the blocks are dirty and each write requires a block to be read, modified, erased, and rewritten on the same block.

 

[Ao1]

128x+ write amplification is real. Life span of 1 year is real.

My numbers indicate I have a write amplification of 175x.

3.2TB of lifetime data written with average P/E burn out of 4498 x 128GB.
562TB / 3.2TB = 175x write amplification

The only thing I do is browse the web and leave my computer on 24/7. Windows system logs, file system journaling, page file writes and browser cache are constantly writing small random data to the drive. I have plenty of ram and free space on the drive. All the small writes will kill your drive. I have replaced the drive recently as Windows finally started crashing even after I reinstalled. TRIM, garbage collection, and compression won't save SSDs.

I agree that a slow drip of writes consisting of small data xfer sizes is really bad for WA, but a WA factor of 175 is extreme. If the sector size is 512 bytes the avg data xfer size is 18 KiB, which seems realistic although perhaps a touch high for the usage you describe. It does however discount any chance that the sector size is based on 4 KiB or 8 KiB blocks.

C6 – 7037727161 x 512 = 3603316306432 bytes or 3.2 TB………. WA = 562/3.2 = 175
C9 Number of write commands – 198005481
Assume 512 byte sector = 3603316306432/ 198005481…….Avg data xfer size = ~ 18 KiB

The wear levelling (WL) efficiency does not look very good. This should be as close to 1 as possible but you have a minimum erase count of 3,444 and a maximum erase count of 5,342 (1.55)

Available writes 128 x 5,000 = 640000 GB
Available writes / WA * WL = Life time expectancy
640000/ 175 * 1.55
640000/271 = 2361 GB = 2.3 TB

Intel claimed that the X25-M's WA at less than 1.1, whilst also claiming that competitors WA was in the region of 3. They also claimed that WL efficiency should be a factor of 1.04 with mediocre SSD’s having a WL efficiency of 5. Since the days of the X25-M most SSD’s now use large DRAM caches to improve WA/WL, however the DRAM cache has to be flushed periodically so if writes are small and infrequent it would be reasonable to assume that the benefit would be negated.

This is where the testing on XS (which I was part of early on) is not representative of normal use as the writes are continuous except for a small interval between cycles. The workload is predominantly sequential and the random element of the workload is most likely write combined by the SSD controller. In other words the endurance workload is a best case scenario.

The main myth that the XS thread has dispelled is that SSD’s gracefully transition to read only once the P/E count has expired. That has not occurred on a single drive that has been tested.

The JEDEC spec states that if you've only used 10% of the lifespan of your device (cycles or GB written), then your data needs to remain intact for 10 years. If you've used 100% of available cycles, then your data needs to remain intact for 1 year.
Testing on the XS thread has revealed that removing power once the P/E count has expired can result in the SSD failing within minutes, so it’s not much use getting excited about how much data they have written past the theoretical P/E count when the integrity of the data is highly volatile. The key factor of SSD longevity has to be a combination of the duration of data retention and the amount of P/E cycles consumed. Only then can SSD endurance be properly understood and despite the testing at XS this is still a completely unknown factor.

Sadly very few SSD’s now report SMART data that enables host writes vs NAND writes to be calculated, which leaves only guess work. For early SF drives (which did report this data) WA was coming out from between 0.58 and 3 depending on work load. (WA below 1 is of course due to compression).

 

[Ao1]

.

The thread over at XT also dispels the myth that any other NAND other than the stuff that Intel and Micron uses on their (own) SSDs is crap, as the thread clearly shows the so called "2nd tier NAND" is every bit as durable as the stuff Intel and Micron are using on their own SSDs.

Really? How come not a single OCZ SSD has managed to get to the end of the theoretical P/E count? The rebranded X25-V on the other hand seems to be doing rather well. Can you back up you statement with a comparison of the drives that have been tested?

If you have a NAND fab facility you can be sure that premium NAND is going to end up in your SSD products and not someone else's.

 

[=Wendy=]

Really? How come not a single OCZ SSD has managed to get to the end of the theoretical P/E count? The rebranded X25-V on the other hand seems to be doing rather well. Can you back up you statement with a comparison of the drives that have been tested?

If you have a NAND fab facility you can be sure that premium NAND is going to end up in your SSD products and not someone else's.

I made an observation based on what was known at the time. My post was made in Nov 2011. Perhaps you point me in the direction of an SSD that had actually failed by that time?

 

[Ao1]

Sure, just check out the first post in the endurance thread, which was last updated 06-27-2012. There are numerous graphs in that first post summarising data that was collected up to that point.

http://www.xtremesystems.org/forums...25nm-Vs-34nm&p=4850357&viewfull=1#post4850357

Can you link me to the info that supports your assertion that SSD vendors using "2nd tier" NAND are doing as well as SSD vendors that make their own NAND?

 

[Hellhammer]

Your Patriot Torx SSD has an old gen Indilinx controller, which is somehow able to get write amplification to horrendously high levels. I've seen several older Indilinx-based OCZ SSDs with similarly high WA values. Wear leveling on such drives must not be very sophisticated.

Modern SSD controllers with light usage typically have a much lower write amplification than that. On good drives 1.5-2.5x is the norm nowadays, not 175x.

The earliest Indilinx controllers did block level mapping, which means that the smallest write you could do was the size of the block (512KB back then). That significantly increases WA because you may have to write the whole 512KB just to write 4KB of data (128x WA). Combine that with poor GC and you have something that's not a winner. Nowadays all drives do page level mapping so WA of 1-2x is realistic for light usage.

 

[Revolution 11]

I still say bit errors of the kind being referred to in that "article" have little to nothing to do with write limits. When NAND wears out due to write limits, it becomes read-only. I don't think a data-recovery company is going to be called in to deal with a NAND device that has switched over to read-only... Those devices are defective. That's a completely different situation.

Well, according to that endurance thread at Xtreme Storage, the problem testers have found with SSDs is that they "don't" fail into a read-only mode.

Originally Posted by devsk
This is troubling, very troubling. None of the SSDs so far have gone into read-only mode upon failure. Not even the latest incarnation as of 2013. This was touted as one of the main features of SSD.

In fact, the failure mode is hard. There is no way to get the data back and programs hang hard. Nostalgically looking, HDDs were way better in this area. I would gladly buy even a slower and slightly more expensive SSD drive which always fails into a read-only mode, leaving the data accessible.
​

That combined with Ao1's observation about the interactions between P/E cycle usage and retention of data is worrying, to be frank.

 

[Hellhammer]

That combined with Ao1's observation about the interactions between P/E cycle usage and retention of data is worrying, to be frank.

It's not worrying when you look at the big picture. The guys over at XS are going over the specified endurance rating and the drive is not covered by warranty when that point is reached. Basically the manufacturer is telling you to replace the drive because its endurance limit has been reached and it may no longer store data reliably. You will not face any of the issues at that point because the NAND will still hold the data for at least one year and the drive is still accessible.

 

[JellyRoll]

I love necros!

 

[Ao1]

It's not worrying when you look at the big picture. The guys over at XS are going over the specified endurance rating and the drive is not covered by warranty when that point is reached. Basically the manufacturer is telling you to replace the drive because its endurance limit has been reached and it may no longer store data reliably. You will not face any of the issues at that point because the NAND will still hold the data for at least one year and the drive is still accessible.

I agree it’s not worrying, I’m just trying to point out that an end user should expect the max amount of writes that the theoretical P/E count can accommodate and no more. That is of course still a lot of writes, which are more than adequate for a typical client application.

Personally I would prefer it if SSD’s converted to a read only mode once the theoretical P/E is reached and that they are then capable of retaining data for at least 1 year with no power once they get to that point.

 

[taltamir]

128x+ write amplification is real. Life span of 1 year is real.

My numbers indicate I have a write amplification of 175x.

As others have pointed out you were using a bad ancient controller...

What you quoted wasn't me saying this is impossible in such drive. But raising the possibility that such horrendous results could ALSO hypothetically occur on a modern drive IF you turn off trim and spare area.

In fact, the failure mode is hard. There is no way to get the data back and programs hang hard. Nostalgically looking, HDDs were way better in this area. I would gladly buy even a slower and slightly more expensive SSD drive which always fails into a read-only mode, leaving the data accessible.[/I]
[/INDENT]That combined with Ao1's observation about the interactions between P/E cycle usage and retention of data is worrying, to be frank.

How was this tested? A program and an OS running from a drive WILL hang if its read only and it is expecting read/write.
The only way to prevent such crashes is to mount it as a secondary drive in a read only mode and copy off data WITHOUT actually EXECUTING any programs on it.