SSD Write Endurance 25nm Vs 34nm

[groberts101]

If a SF drive is throttled or not depends on the settings the vendor selects. That is equally true for the SF-2281.

According to Mushkin:

"DuraClass management functionality is still active. “Unthrottled” in this context refers to write IOPS bursting up to 90,000+ but being governed down to 20,000 after a few seconds which is typical behavior with standard firmware with SF-2281. The firmware we have on the Chronos and Chronos Deluxe drives will not have that governor activated."
the term "unthrottled" in that context has only to do with short burst rates. NOT anything at all to do with sustained incompressible write speeds which is one of the only true ways to really see throttles on Sandforce controlled drives. For example.. you can test a drive with ATTO before and during throttling and the posted speeds will rarely change more than about 5%. Use incompressible data and the picture changes dramatically as the drive cannot leverage it's internal compression algorithms to falsify/fudge the numbers.

For simplicities sake.. there are 3 distinct stages to Sandforce's speeds.

1'st is "unthrottled state"/fresh speed and is maintained until a map of the entire physical space is formed(all available physical space written to at least once).

2'nd is called a "settled state" and will only write incompressible data at about 70% of its original fresh speeds(this is the "lifetime throttle" and will in fact be stopped immediately with secure erasure as the Durawrite process requires an actively mapped state of the entire physical space for it to be implemented. You can idle for days straight and this throttle will never be lifted(hence the term.. "Lifetime Throttle"). Get rid of the map with an SE?.. and Durawrite has no way of knowing that the drive should be throttled without it.

3'rd is called a "hammered state" and is implemented when the drive is written too hard and all fresh blocks have been completely depleted. At this phase of operation the controller will ony maintain about 50% of it's original fresh speeds when writing incompressible data. This is where the idle time comes into play as the controller rebuilds its fresh block reserve and que's the controller to allow the throttle to eventually lift. This is the ONLY state(although there is some grey area in here) that idle "speed recovery"(NOT talking about block recovery) is possible. And trust me here... with a 6 drive array?.. I know when I lose 110MB/s incompressible write speeds with my vid work not to mention that I immediately notice the small lags that stem from the increased latency of this algorithm(especially with heavy multitasking).


Excluding the burst speed throttling there is also a different type of throttling implementation that brings the drive to its knees if you write "excessively" for a sustained period. Yeah.. see above. Is called a "Hammered State".

The only way of alleviating that is idle time, which is completely retarded, as it has no relevance to the date on the receipt. Correct.. idle time can recover speed ONLY back to a top mark "settled state" and the lifetime throttle remains in effect.

SE will not clear this. That data is held in a secure area of the drive. You can SE your drive once or a hundred times. It won't clear it. Completely incorrect. I've purposely hammered my drives by filling them many, many times beyond capacity(once even wrote over 1TB per a single degradation test.. yeah I know.. I've heard it all before) and I know the internal algorithms/throttled stages of these drive better than most and have schooled many "experts" by now.

I would suggest that if you are able to restore performance via a SE the drive was most likely in a degraded (not throttled) state. As mentioned above there are various stages of throttles with these controllers. You can recover form ALL throttles with an SE.

Not to toot my own horn here.. BUT.. I actually tested these drives more than many others combined(conservative estimate of 10 Petabytes worth of data written for highly specific degradation/recovery testing)) and there are fully formed current maintenance protocols based on what I and others who did extensive testing found in regards to these internal algorithms. All the info was spread from the stickies/threads at OCZ forums and are now in othe mfgrs forums as well.

I know it sounds y, condescending and "know-it-all" like(I hate that too but in the end?.. I listen to the one's "in the know" as this stuff is about learning and not about pride or ego), and I don't actually really mean for it to be that way since after all.. I'm sharing here, but you would be pretty hard pressed to tell me anything about the sandforce controllers inner working that I don't already know. Is one of the main reason s that I was chosen to be a beta-tester for them.

[Ao1]

So how do I clear throttling from the V2 I used in his experiment? I have SE's numerous times and its only idling that does anything.

Also I wrote 30TB in one sitting

SF vendors should come out and make a clear statement on exactly what throttling measures have been applied and what the impact will be.
It should be a clear specification on the product.

[bulanula]

SF vendors should come out and make a clear statement on exactly what throttling measures have been applied and what the impact will be.
It should be a clear specification on the product.

Would like to see this as well for the sake of transparency !

[Computurd]

[Note: I find your rather unusual method of trying to quote someone by adding your own red colored text to the quote extremely confusing and difficult to read. I had to check back a couple of times to make sure that Ao1 himself had not actually written what you wrote. There is a quote tag for a reason you know. Even italics would be easier to read than text color changes.]

actually i quite like it. and might even steal the method and use it myself

@AO1...hey i recall that....seems that we based a good drive temp from that document to be 30-40C , i still set the curve on my fans on my HDD to never allow them to go over 40C because of that document. Even though the conclusion was that it didnt have a huge impact on performance, you can see in figure 4 that the best range for lowest failure is 30-40C. that is still widely held knowledge. google does us a great service releasing data like this!

[johnw]

actually i quite like it. and might even steal the method and use it myself

I did not even realize he had inserted his reply in that long quote until someone pointed it out.

So feel free to use the technique. I'll probably miss your replies, too!

By the way, I'd say the safe range is more like 25-50C from that graph, although the error bars are large over 45C, so it might be better to set the upper limit at 45C, then to make it almost symmetrical, the lower limit would need to be like 30C. So 30-45C.