Sandforce Life Time Throttling

[Ao1]

You obviously haven’t been following the endurance thread. So far out of the 4 different SF vender drives that have been tested only OCZ drives have been throttled.

Intel specify what work load is covered by the 5 year warranty. If you exceed that by an excessive margin you void the warranty.

Regardless of how unlikely it is that someone could induce throttling the word "enthusiast" and "throttling" don’t mix.

What is the proprietary Indilinx Ndurance Technology by the way? Another form of throttling writes? If an Octane was put in the endurance thread will it become throttled at some stage?

[Christopher]

You obviously haven’t been following the endurance thread. So far out of the 4 different SF vender drives that have been tested only OCZ drives have been throttled.

Intel specify what work load is covered by the 5 year warranty. If you exceed that by an excessive margin you void the warranty.

Regardless of how unlikely it is that someone could induce throttling the word "enthusiast" and "throttling" don’t mix.

What is the proprietary Indilinx Ndurance Technology by the way? Another form of throttling writes? If an Octane was put in the endurance thread will it become throttled at some stage?

The way I understand the 520 as far as warranty is concerned is this:

Retail box units get a blanket 5 year warranty.
OEM units get 5 years or until MWI = 0, which ever comes first.

Previously, if you had purchased a 320 OEM, you'd just get one year. But that was the only way to get 7mm drives (OEM boxed units). Its good to see they've changed it.



No SF vendor, except OCZ, uses LTT as far as I know. No one is using throttled IOPS anymore either. I don't know whether the OCZ 2281s are still using it, but I don't think anyone wanted to gamble on that.

[Ao1]

Intel’s warranty is a “limited” warranty. Limited generally implies that the warranty is subject to limitations, conditions and exclusions.

Intel publish how much a client SSD can write per day for the duration of the warranty. If you exceed that amount before the end of the warranty period your warranty will not be valid.


OCZ’s warranty is also limited. Despite LLT:

Limitation of Warranty:

• Connection to a faulty power source
• Alteration, Modification, Disassembly or unauthorized repair
• Improper use of product
• Normal wear and tear
• User inflicted intentional or accidental damage
• Any other cause not resulting from defects in materials or workmanship

[Christopher]

Ao1 -- You're right, of course. But I doubt Intel is going to check anything aside from the purchase date at their RMA center for a drive with a retail box SKU.

From a 520 OEM product listing:

5 years limited

All Intel SSD 520 products sold in the OEM boxes have a five-year limited warranty term; however, the warranty includes a "Media Wearout" limitation. The Media Wearout limitation provides that the warranty shall expire when the usage of the drive has reached a predetermined usage limit established by Intel, which could result in a warranty term much shorter than five years for drives used in heavy-use, enterprise applications. The Media Wearout is determined by Intel's implementation of the SMART attribute "E9" Media Wearout indicator (as measured by and shown in the Intel SSD Toolbox, www.intel.com/go/ssdtoolbox). At any point during the warranty term, if the Media Wearout Indicator reads "1", as measured by the Intel SSD Toolbox, then the drive has reached its wear limit and the Limited Warranty at that point expires, regardless of how much time may have been remaining on the five-year term. This Media Wearout limitation is imposed due to the greater workload demands of enterprise/data center deployment on SSDs

That's a lot better than previous OEM boxes. But I think the retail boxes are effectively "if it dies before 5 years, you get a new one"

I only checked the 520's OEM warranty because I got hosed when I ordered a 320 and ended up with an OEM box drive with 1 year of coverage.

[Ao1]

Samsung 830

1st a caveat. It is not known for sure if SMART attribute 177 is counting P/E cycles. If that assumption is wrong the output is wrong.

The Samsumg drive has 10GiB free out of 59.5GiB. I started to monitor as soon as I had finished installing the OS and apps. I then subsequently used the SSD as I normally would.

The first change in the P/E count came as quite a shock. 9GiB of host writes vs 60GiB of NAND writes.

This really had be thinking that SMART attribute 177 was not counting P/E cycles, however subsequently I have been able to write 22GiB of host writes without a change in SMART attribute 177. So what the beep? I’m going to guess that the huge amount of NAND writes was due to GC following the installation of the OS and apps.

I’m going to wait for the P/E count to get to 15 to get a better average and then I’m going to do some tests with both the 520 & 830 as non OS drives. I’m going to SE both drives and fill them with 4K random writes. I will then stop all write activity and see if the NAND writes increase as the drive tries to clean up the mess.

Also in a couple of weeks when I’ve finished the above I’m going to get a Petrol drive to see if the “proprietary Indilinx Ndurance Technology” includes a life time throttle feature. (Assuming OCZ do not reply to my query over in their forum on this issue).

[Christopher]

Ao1,

Both of your 64GB drives came on 01FW, right? Remember, my 64GB 830 didn't count 177 correctly, so there is no telling how many permutations of that issue exist. The change to 02FW didn't really fix it as much as make it less retarded. I am certain it is a PE cycle counter, but I'm not certain that it works correctly (all the time).

I've got one 830 still working on 177 = 0, and another one which flipped over to 177 = 1 at 140GB (on a 256). I think it would be very hard to look at a one PE cycle difference to discern much -- depending on how it calculates the number. I always assumed it was analogous to the Indilinx Avg PE count.

[Ao1]

^ SMART attribute 177 increased yesterday to 14 giving a more respectable WA figure, but I’m still not 100% convinced that it is recording P/E cycles. Samsung support stated: “The raw value on this attribute is how many times it can prolong the life of a specific block” , which might explain why it could appear erratic.
It might refer to something like a write combining event that was able to prevented a block erase. I’ll continue to monitor until it gets to 15 to see how things progress, but somehow I don't think SMART attribute 177 is quite figured out just yet.

[Ao1]

Back to the 520.

This is a quick rundown to hone onto block sizes and QD’s that result in either good or bad WA. The results are a bit rough and ready due to the small amount of writes for each category and the limitation of 1GiB increments for NAND writes, so results could vary, especially for the small blocks at QD1.
A pattern does seem to immerge though

• 0 fill appears equally good regardless of block size or QD, except 1K
• Incompressible is actually quite impressive, except 1K

So it seems that random writes of 4K or less, regardless of compressibility, have (to a varying degree) a disproportionate impact on WA, which would explain why WA was coming out so high when just monitoring typical OS and app workloads.

[canthearu]

Awesome work Ao1, wish I had the option to thank as well.

I definitely agree, from my own observations over the last week, that very light OS activity can produce bad WA on the sandforce drives. I did get my drives down to 1 or 2 gig per day NAND writes a day though. (24/7, just sitting doing not much)

I don't really think it is a problem though, the usage is so light that it would take decades to wear the drive out. The controller would have panicked for some reason or another long before that

[Christopher]

Well, I bought an Octane, and I'm going to start endurance testing it. It's lack of SMART data is disheartening, much like the Marvell drives.

http://www.xtremesystems.org/forums/...=1#post5072822

So I guess the real question is: If it has throttling, when and how will it kick in? Will it work on a similar [to SF] "credit" system? Is nDurance just a marketing creation? Who knows? Not me. Not yet.

[kensiko]

Someone asked about the nDurance in the forum, but we got no answer. Nothing is really found on google also.

[Christopher]

So the Octane is acting MARVELLOUS so far. Perhaps nDurance will rear it's head soon.

[Ao1]

Here is the final comparison of WA for the 830.




Edit: No sign of a temp throttle. Oh well, it was worth a try.

[johnw]

Here is the final comparison of WA for the 830.

Interesting data!

So, assuming Samsung 830 attribute 177 is telling us what we think, the 830 had WA=2.2 for the cumulative load you tested, while the Intel 520 had WA=1.4

It would seem that the Intel has the advantage. Still, you were only averaging about 8GB per day (flash writes) on the Intel, and maybe 14GB on the Samsung (the date is not visible for the last Samsung column), so either SSD should be able to manage that level for 5 years easily, and perhaps quite a bit longer.

But I've always suspected that the WA on the Samsung is higher than most other SSDs. So I wonder whether the difference you measured is due to compression, or more likely, to just the algorithm Samsung is using resulting in higher WA. I wonder what a similar test would measure for WA on a Crucial m4 or Intel 320. I wouldn't be surprised if the WA for a Crucial m4 or Intel 320 (or C300 or X25-M) were lower than the Intel 520 under the same test conditions as you used.

[Christopher]

Thanks for giving the Hair Dryer Test a go, Ao1.

[Ao1]

Whoops, I missed off the last date. I’ve updated the chart and have added a few more stats. I don’t know why the first WA reading was so high for the 830. Take that out of the equation and the write activity was nearly the same per day for both drives, but WA was marginally lower for the Samsung.

[johnw]

Whoops, I missed off the last date. I’ve updated the chart and have added a few more stats. I don’t know why the first WA reading was so high for the 830. Take that out of the equation and the write activity was nearly the same per day for both drives, but WA was marginally lower for the Samsung.

I don't think the write amplification averaged over time is a useful quantity. What matters to the longevity of the flash is not the time average of WA, but rather the total amount written to flash (neglecting second order effects). So, the cumulative WA is probably the most useful quantity.

For the Samsung, that is either WA=2.2, or if you leave off the first day, then there is 120GiB of flash writes and 72GiB of host writes, so WA=1.7.

Which is compared to cumulative WA=1.4 for the Intel 520.

I still predict that the same load would result in an even lower WA for an m4 or 320.

[Christopher]

Ao1, your 830 is on 03FW, right?

[Ao1]

Hmmm, after posting yesterday I hit the 830 with a lot of 4MiB xfers. WA 0.95.

F/W Version is CXM03B1Q

[johnw]

Hmmm, after posting yesterday I hit the 830 with a lot of 4MiB xfers. WA 0.95.

But that was only one increment of attribute 177, right?

I'm not sure how accurate it is on that scale. The cumulative WA is 1.7 now, and that goes from sa177=12 to sa177=16, so it should be more accurate.

[OldChap]

Bought an Intel 520 120GB (the retail version) and have it running on a Linux Mint box. There seems to be not love for Linux by Intel when it comes to the toolbox so I'm looking for advice on a linux equivalent.

I tend to download around 10TB pcm running MJ-12 plus other compression processes but this is not set up as an OS drive. I have over provisioned to leave 96GB (I hope I did that right).

I have some utility that only reports attributes 170, 184, 187, 232 but nothing else

Thoughts anyone. ????

[canthearu]

Bought an Intel 520 120GB (the retail version) and have it running on a Linux Mint box. There seems to be not love for Linux by Intel when it comes to the toolbox so I'm looking for advice on a linux equivalent.

The 520 doesn't really need the toolbox. It only receives about 10% improvement from using the optimizer under windows anyway. Edit: and this improvement is only until the cleared NAND is filled again, and then it is back to steady state speed.

I tend to download around 10TB pcm running MJ-12 plus other compression processes but this is not set up as an OS drive. I have over provisioned to leave 96GB (I hope I did that right).

You should use the SSD under linux just as you would use it under windows 7. Just ensure ext4 is the filesystem you use, and discard is added to the drives options in /etc/fstab. This will enable TRIM support. You do not need to generally set up over-provisioning when you have TRIM.

I have some utility that only reports attributes 170, 184, 187, 232 but nothing else

Thoughts anyone. ????

Try:

sudo smartctl -a /dev/sda

to list all smart attributes. Use intel's manual for the 520 if the attribute names don't come out right.

[OldChap]

Thanks I will look at smartctl in the morning.