Cherryville - SSD 520

[Reflex1]

Im looking to upgrade my X25-M ...is this the drive to go for? A vertex 3 Max Iops is the same price where i am. Performance is my first priority, reliability second

[Anvil]

I have no issues with recommending the SF drives, they've been great on my systems.

In real life there is really no difference running a standard Vertex 3 vs the MaxIO.

Anyways, you didn't mention price (nor capacity) as a parameter but as reliability is second then just go for the cheapest Synchronous drive. (i.e. Vertex 3 or the Samsung 830 or the m4 )

--

I've just started testing the Intel 520 180GB and so far it feels like any other ~240GB drive.

I'll do some tests during this weekend using hIOmon just to check how TRIM behaves vs the standard SF firmware and then try to compare to the V3 standard and possibly either the MaxIO or the WildFire (using Toggle mode).

[Anvil]

F9 is equivalent to E9 on the standard SF firmware (Raw NAND writes) and is listed in GiB.

F1 (and E1) Host writes and F2 Host Reads are not listed using GiB, I haven't checked yet but it could look like it's increased every 32MiB.

HD Tune is one of the few that can display SMART info on the X79 chipset so the usual CDI can not be used.
I'll check if HDSentinel works.

10-February-2012_22-19.png

[Computurd]

Here some bench with SNIA:

LOL they have nothing in common with the SNIA standard.

there is no 2x writing of the device...there is a write-fest (with workload independent preconditioning, so a total different type of load than what you are testing) that culminates when the drives performance does not fluctuate +/- a valuation: then you test with the actual workload yoiu are testing.

2.1.21 Steady State: A device is said to be in Steady State when, for the dependent variable (y) being tracked:
a) Range(y) is less than 20% of Ave(y): Max(y)-Min(y) within the Measurement Window is no more than 20% of the Ave(y) within the Measurement Window; and
b) Slope(y) is less than 10%: Max(y)-Min(y), where Max(y) and Min(y) are the maximum and minimum values on the best linear curve fit of the y-values within the Measurement Window, is within 10% of Ave(y) value within the Measurement Window. that is how you reach steady state, then you log. there is also no wait time, then resumption of testing.

i see no similarities to SNIA spec.
even with in-depth talks with the head engineer from Calypso, and members of SNIA specification committee, there really is no way for an end user to perform full SNIA spec testing without dedicated hardware. the time and effort would be absolutely ridiculous. you have to re-prep for steady state for each and every single test specification ran. hours upon hours upon hours, even when done automatically. oh well.
its a great concept, but horrible application. Micron/others are very interested in getting the review community using these types of standardized specs, but without a $25,000 machine, you can forget it. oh and that is the price they quoted me with a "generous" discount!!!

[Computurd]

and did anyone else notice how awesome that M4 did in that "SNIA" test? wow!

im not saying that test does not have worth, btw. it does. they just shouldnt be claiming any type of SNIA compliance....

[Ao1]

Micron/others are very interested in getting the review community using these types of standardized specs, but without a $25,000 machine, you can forget it. oh and that is the price they quoted me with a "generous" discount!!!

Why don’t they publish their performance specs based on the SNIA spec?

I’m disappointed Intel did not use SNIA to establish product specs. Instead their product sheet states that the specs are based on “out of the box” performance (for random 4K write). They also state specs are based on Iometer, but they do not mention they used a highly compressible data setting.

[Computurd]

well there is a 'push' in that direction amongst several of the manufacturers. But you would run into that same old issue, other companies wouldnt do it! Joe Blow is still gonna buy the SSD with '800 MB/S!!!!!' over one that is marketed with a '250 MB/s SNIA Verified' spec. thats just the nature of the beast.

I would look for it to begin to be used for Enterprise-class devices soon, where the buyers are definitely going to be educated about the SNIA spec and what it actually means.

but for consumer, well...dont hold your breath!

i suggest using two forms of specs for consumer devices. But...one must also ask, do you think a consumer is going to get these drives into these wild steady-state conditions? with trim and GC, the playing field right now is very fluid...its like you are giving the specs on two ends of the spectrum, where does the true consumer usage lie?
SNIA has tried addressing this concern somewhat, with two different sides to their testing, client and enterprise. I personally do not feel that they go far enough...yet.

[iboomalot]

I would say the price is going to keep out alot of consumer people.

The 520 480 gig is $1,049
The 830 -512 is $799
The M4 - 512 is $654

What little extra speed the 520 has for the consumer won't justify that extra cash outlay and a reduction of 30GB of storage space.

1-2 secs faster boot to windows won't offset $400 more cashflow

but does make for interesting convo

[Ao1]

well there is a 'push' in that direction amongst several of the manufacturers. But you would run into that same old issue, other companies wouldnt do it! Joe Blow is still gonna buy the SSD with '800 MB/S!!!!!'

I can understand marketing pressure, but I expected more from Intel. What next, they bring out the next SF controller without any validation just to get to market at the same time as other vendors that do not bother with validation?

They were actively involved in developing the client performance test. The performance test recognises that FOB is not representative, yet that is exactly what Intel has specified for the 520.

Whilst the test is based on non-compressible data there is nothing to stop a vendor using compressible data as long as they state clearly that compressible data has been used. The test also requires all volatile caches to be disabled, which should be an advantage for an SF drive.

I do not trust the review sites that make up their own synthetic benchmarks, when it is abundantly obvious that in some cases they have been developed to make certain SSD shine rather than reflect true performance in a client environment. It’s quite common to see vendor A being directly compared to vendor B when vendors A’s drive is double the capacity. The simple way to avoid confusion is for vendors to specify performance using SNIA.

[squick3n]

How do these drives perform in a non-TRIM environment, compared to something like the Samsung 830 or Crucial M4? It really seems like the effectiveness of GC varies from review to review

[Anvil]

@CT and Ao1

I do agree on a lot of things defined by SNIA but I also can see that it's more or less impossible to stick 100% to the testing methods defined, it's just not feasible.

The issue here is that drives are not preconditioned equally, that is, in most cases they sort of are as review sites and most users often post benchmarks showing best case results.
Even if one was able to create a workable subset of the tests it would be a time consuming test as preconditioning still would have to be done prior to each test, the question I ask myself is, is it really worth it?
From a testing point of view it is much more convenient to secure erase a drive between each test as it literally takes just a few seconds, it will show best case results but the results will be reproducible, in most other cases you cannot reproduce test results.
From a client point of view the SSD will in most cases be used in a TRIM environment and that will in most cases result in performance that is closer to a secure erased state than a degraded state.

Last night I performed some tests on the Samsung 830s in raid-0 and GC is really kicking in, it is clearly doing what it's supposed to, so having a very active GC can make such testing invalid as well as the drive will continuously try to "recover" from a degraded state. It did not take long for the Samsung to clean up, we are talking minutes not hours.

[Anvil]

How do these drives perform in a non-TRIM environment, compared to something like the Samsung 830 or Crucial M4? It really seems like the effectiveness of GC varies from review to review

The Samsung 830 using the latest fw looks to be doing really well in a non-TRIM environment.
The Samsung is doing active GC while the m4 cleans up when needed, from an endurance point of view the latter yields more Host Writes as active GC in general will end up using more cycles.

In any case the Samsung will most likely outlast the rest of your setup and so will the m4's.

Let's stick to 520 related info and or testing in this thread

[squick3n]

nm. I misunderstood. Thanks for the response Anvil

[Ao1]

@CT and Ao1

I do agree on a lot of things defined by SNIA but I also can see that it's more or less impossible to stick 100% to the testing methods defined, it's just not feasible.....

I can see that the time/ cost involved means that SNIA is not going to be a favourite with review sites, but I do think that vendors should specify their specs using the test. That does not mean that they can’t use other benchmark assessments as well, but at least there would be one common and indisputable test method. [EDIT: All major vendors contributed to developing the test]. Right now some vendors use different benchmarks for different performance aspects to cherry pick the best performance figures possible.
Review sites should stick to real world tests or be a lot more specific about their synthetic testing.

How did the trim exercise come out? I can’t help think the only performance difference with the 520 is related to the fact that RAISE is disabled, but it would be interesting to see if the TRIM “hang” exists with the 520.

EDIT:
And the problem with FOB tests is that some controllers have much higher overhead once in a steady state. FOB performance does not last long once the SSD is used.

[therat]

Does anyone know if the 520 SSDs have throttling enabled?

[Ao1]

^ Great question. If they were not so expensive I would buy one to find out.

[some_one]

AFAIK, 520 SMART values...

[Computurd]

I do not trust the review sites that make up their own synthetic benchmarks, when it is abundantly obvious that in some cases they have been developed to make certain SSD shine rather than reflect true performance in a client environment. It’s quite common to see vendor A being directly compared to vendor B when vendors A’s drive is double the capacity. The simple way to avoid confusion is for vendors to specify performance using SNIA.

absolutely. i couldnt agree more. even outside of the capacity differences. capacity comparisons is an issue, as there is simply a limited supply of review units that go out. companies choose what capacity, and thats what you get. some send higher, some send lower. some want them back, some do not. unfortunately you are dealing with the 'whims' of the mfg in that aspect.

there is a huge need for standardized traces. I can make a trace right now that would make any drive look better than another. its just a simple means of 'tweaking'. i think right now if there is one area that is totally 'out of bounds', it is site-specific trace based testing. everyone is growing their own builds, and most have no worth.
unless they were used correctly of course, as a comparison of FOB v steady state performance of the respective trace. even that would still be open to abuse, but at least would provide some comparative data. maybe this would be better, but still not perfect.

[Christopher]

While I agree to an extent, I don't think sites were making up their own traces just to make SandForce look good (if that is what is really at the heart of this). It doesn't really matter to me one way or the other as I don't buy what is fastest; I buy what I think is most interesting, and that's not something that any trace is going to reveal.

Fortunately, there are a lot of interesting drives.

[Anvil]

How did the trim exercise come out? I can’t help think the only performance difference with the 520 is related to the fact that RAISE is disabled, but it would be interesting to see if the TRIM “hang” exists with the 520.

EDIT:
And the problem with FOB tests is that some controllers have much higher overhead once in a steady state. FOB performance does not last long once the SSD is used.

I haven't had the time to install hIOmon yet, will do later this week.

FOB performance can last very long (for some controllers ) and is highly dependent on usage pattern/environment,...

[Ao1]

Intel has now posted the product specs for the 520, including a piece about compression.

http://www.intel.com/content/www/us/...ech-brief.html

I don’t want to be unfair to SF and maybe I am just not getting it, but it seems to me that 0fill does not happen outside of synthetic benchmarks. It’s like a car manufacturer giving miles per gallon performance for driving a car on the moon. It’s just plain wrong.

Once the SF controller has to deal with anything that that can’t be compressed by more than 75% performance drops like a stone for both reads and writes. For a client based SSD let’s say everything could be compressed by 75%. If you compared performance benchmarks using data that can be compressed by 75% the SF would suck compared to other current offerings and in reality most data can’t be compressed by that amount.

Am I missing something?

[minpayne]

Has Intel addressed the BSOD issues of the SandForce controller? At launch time it seems I could only find a single case on the Intel forum.

If it's almost BSOD-free I'd start considering this drive.

[Anvil]

Well, an Intel 6Gb/s controller is available
Not much on goolge but here is the model : EW29AA31AA1

If you buy the Hitachi SSD you can get one now

Link to StorageReview

[johnw]

Intel has now posted the product specs for the 520, including a piece about compression.

http://www.intel.com/content/www/us/...ech-brief.html

I don’t want to be unfair to SF and maybe I am just not getting it, but it seems to me that 0fill does not happen outside of synthetic benchmarks. It’s like a car manufacturer giving miles per gallon performance for driving a car on the moon. It’s just plain wrong.

Thanks for the link, I hadn't seen that yet.

I can think of three POTENTIAL benefits to compression by an SSD controller:

(1) Higher performance and endurance by having essentially larger overprovisioning due to stored data being compressed

(2) Higher performance by reducing flash writes

(3) Higher endurance due to reduced flash writes


Of those three, I think only (1) is likely to be significant for most users. It seems that most Windows OS installs can be significantly compressed by the Sandforce controller. If the Windows install takes up 20GB, then there may be as much as 5-10GB of flash writes saved, which (if the firmware takes advantage of it) should have the same effect as reserving an extra 5-10GB of flash for overprovisioning.

I don't think (2) is significant, because as you say, most data users write to their SSD (after initial OS and application installs), will not be compressed significantly. I am suprised that Intel claims compression to 85% of original size for Microsoft (I assume Office) documents. The Microsoft Office documents I am familiar with are automatically compressed (Word, Excel, PowerPoint) before being written to storage. I would be VERY surprised if Sandforce can compress such documents at all, let alone to 85% of their original size.

For the same reason, I think (3) is not going to be significant.

The only possible exception I can think of for mainstream usage that might make (2) or (3) significant is for someone who uses hibernation frequently (several times a day) and has an easily compressible hibernation file of 4GB or more written to SSD frequently. But I am just guessing here, since I don't use hibernation and I have never examined the hibernation files carefully. Does Windows do any simple compression on the hibernation file before writing it to storage? If it does, then it is unlikely the Sandforce controller could obtain any additional compression on it.

[Ao1]

All good points. I know that a Windows and Office install can be compressed down to 50%, so straight off the bat more or less every drive sold for a client application will have ~4GB of NAND for the controller to use as it sees fit. Despite the benefit of compression however none of the SF drives have performed any better than non SF drives over in the endurance thread, even when using 46% compression. The MWI does not seem to last any longer and the drives seem to burn out a lot quicker once the MWI has expired.

Oh and I was also a bit surprised by some of Intel’s claims regarding compression. Even if they were true however not one of the tasks that they identified would benefit from faster read/ write speeds.