Sony Z11 - Long Term SSD Performance - Post your results

Introduction

Greetings,

There has been tons of debate on the merits of Sony's new dual, tri and quad-SSD configurations in the second generation Sony VAIO Z11 laptops. Sony's decisions on the drive's size, the implications that RAID 0 has on reliability, the drive's physical dimentions, the storage configuration arrangements and total storage size has caused a flurry of posts and caused much confusion.

For the uninitiated, Sony's new Z11 series has been equiped with several variations of a Sony-designed, sub-1.8" SSD that is comprised of Samsung NAND chips and controllers. These drives are exclusively configured in RAID 0 arrays in models with 2, 3 and 4 drive configurations. Sony is the first and only company to ship a consumer laptop product with this configuration.

EDIT 3/28/2010:
Sunfox has compiled a list of the SSD drives and their specific configurations:

My main concern, and the reason for this thread, is the fact that ALL current generation SSD drives have an intrinsic flaw; they loose performance over time. Now granted, storage and software companies have implemented mechanisms and procedures that mitigate the problem to a high degree. The problem here is that nobody truely knows how this is done on Sony's newest laptops and what impact it has on SSD performance over time.

I want to know the following:

1. Do the new Sony SSD's become slower over time?
2. Is there a mechanism being used to prevent performance degradation?
3. If there is a slowdown, how much and how quickly?
4. What does this mean to user's long-term experience?
5. What can users do on their end to maintain the drives for peak performance?

Disclaimer:
I'm not proclaiming myself to be an expert in storage technology, flash memory design or in anything related to this. I'm sure some of the things I've said or am going to say are false. I'm just a computer enthusiast with quite a bad case gizmo addiction. I find this stuff fun.

I'm looking to this community support. Please correct me where I'm wrong and give me more information to add. I'll keep these core posts up to date.
________________________________________________________________


SSD Drives vs Traditional Drives

Now, on to the issue at hand. Why are SSD's different? What is garbage collection? Why is it necessary?

Let's let Intel start us off...

To summarize, SSD's have an extra requirement when writing data that magnetic systems do not. It first needs to read the block before it can determine if it can be erased, and then written (read-modify-write). This is easy when the drive is new and mostly empty. It gets harder to do as files are added and deleted.

NBR Forum Member Psyang wrote a fantastic analology to describe this behaviour and introduces us to the concept of garbage collection.

So the key to sustained SSD performance is the proper maintenance of your marbles (sorry Peter). This maintenance is called garbage collection. Garbage collection is the act of seeking out eraseable blocks that were not processed when the data contained within them was moved or deleted. THis is, of course, an very intensive operation best accomplished when the drive is idle but SSD's have a lot of work to do.

There are two other factors that need to added to this ongoing maintenance nightmare and are the ones I know the least about.

1. Erase block size - They range from 128K to 1 megabyte. This means a 4k single write can force the SSD write up to 40x more data to fill the block. This is known as write amplificication. This also has an impact on how partitions should be aligned.
2. Wear Leveling - NAND memory has millions of individually erasable segments. Each of these segments can be used a limited number of times before becoming unreliable. These are called erase cycles and vary from 1,000 or over 100,000 cycles. Wear levelling allows a drive to rearrange its data so the writes are not concentrated in one area, prolonging the life of the entire drive. Some SSDs will move data around to exercise all of the blocks in the drive more evenly.

Hopefully this has explained why SSD's do what they do to function.
________________________________________________________________


The Introduction of TRIM

The next question is why is garbage collection even necessary? Shouldn't the SSD know the state of each block, making the extra work unneccessary?

This sounds like an easy question until you understand that until very recently, there was no mechanism to update the SSD's mapping of which blocks are used or unerased.

Since traditional HDD's didn't require the extra command to mark a block as eraseable, there was no ATA command necessary for it to complete it's disk activities. ATA Commands are low-level instructions that are part of the "PC/AT Attachment" standard used or emulated by all hard drives.

This is, in a nutshell, what the advent of the TRIM command was designed to do for SSD drives.

Back to Psyang's analogy:

________________________________________________________________


TRIM and RAID on the Sony Z11

Currently, there isn't a way for TRIM commands to pass through RAID controllers/drivers. This means that even if the SSD drives support the commands, they will not be sent to the drives due to the RAID drivers or BIOS config. Additionally, since there isn't a way to change the Sony BIOS's setting from RAID to IDE or AHCI, configuring the SSD's in JBOD doesn't have an impact. TRIM commands are still not passed to the individual disks.

Rumor has it that this may be fixed in future versions of the Intel RST (v 9.6).

EDIT 3/21/2010: Intel has released the Rapid Storage Technology driver v 9.6. Initial testing is showing positive results.

From the Intel help file:


EDIT 3/28/2010: There have been reports that 9.6 DOES NOT ALLOW TRIM.

EDIT 4/1/2010: I'm confused why Intel would have information in the help files of two different install packages released at different times.

EDIT 4/4/2010: Intel has posted that the help files are incorrect. TRIM does not function in RAID with RST 9.6.
http://www.intel.com/support/chipsets/imsm/sb/CS-031491.htm
________________________________________________________________


Performance Testing

Test tools
To collect perfomance information on the Sony SSD drives, we need to look over a common set of tools. I'm open to any other recommendations like WEI or the like.

1. ATTO Disk Benchmark v 2.41 - Default settings
2. HD Tach v3.0.4.0 - Long Bench, Windows XP (SP3) Compatability mode
3. HD Tune Pro 4.01 - Benchmark, default settings
4. HD Tune Pro 4.01 - File benchmark, default settings
5. CrystalMark v2.2 - All tests, default settings
6. CrystalInfo v 3.5.2 - Informational only

Iterative Data Collection Procedure
This is how I'm running my tests 99% of the time. If we are going to compare stats, please try to do the same. I chose the number of passes based on my experience running the tools. CrystalMark, for example, can change dramatically from the first pass to the second. It becomes more predictable after three passes.

1. Reboot device
2. Wait 5 minutes
3. Execute CrystalMark, three full passes
4. Execute ATTO, default settings, two passes
5. Execute HD Tach, default settings, "Long bench" option,two passes
6. Execute HD Tune "Benchmark" tab test, default settings, one pass
7. Execute HD Tune "File Benchmark" tab test, default settings, one pass
8. If a raw partition is available, execute HD Tune "Benchmark" tab test with the "Write" option, default settings, one pass

 

Here is an example of the output from the test tools.

Sony VPCZ112GX/S
RAID 0 - 2x 64GB

 

Intel released version 9.6 of the RST. I've installed it and here is an initial pass.

 

Awesome thread you have created here Zoinks.

What benefits are offered by ver 9.6? Do you recommend the upgrade?

Edit: Oop! never mind. Already answered in the other thread

Also: What magic are you employing to get 30MB/sec on small block writes for a 64x2?
Even The raid 4 folks are only getting 20, and all the other raid 2's are 10.

 

Here is an image I had created when people were just getting their machines and posting them on the original thread:

 

I noticed that too; i think that's related to setting write-back cache to enabled. My Z came with that set to disabled.

also - I can't get crystaldiskinfo to give me a valid report - just get 'disk not found' can't find out what i'm doing wrong yet.....

 

Correct, turn on write-back cache in the RST console.

 

What exactly does enabling write-back cache mean? I tried looking up that particular function and found out that it increases performance at the cost of an increased data loss risk (in case of a power failure). Is it safe to turn it on then? What is the trade-off here?

 

Hi zoinks,

any diff you have noticed after installing 9.6 version? is TRIM supported now?

 

Before turn on write cache back:

After turn it on:

OMG Look at the 4k random write
PS: My Z is running RAID 0 64GBX4

 

There isn't a way to confirm. I'm going to do some more load tests and a TTrim pass.

The Intel RST help file is ambiguous.

 

It allows the controller to cache writes in a way that doesn't immediately write them out. In that way it can aggregate multiple I/O's and minimize the number of write transactions. Thus it has its largest effect on small block writes.

Downside is if power is lost while multiple transactions are in cache, they will be lost.

The good news is we are on laptops so that is unlikely to ever be a significant problem (since laptops have batteries).

 

You mean "Here is an image emev had created ..."

http://forum.notebookreview.com/showpost.php?p=6007733&postcount=967

http://forum.notebookreview.com/showpost.php?p=6008638&postcount=995

 

here's my stab at the benchmarks per the procedure.

VPCZ119GX -- i5 540, 2+2gb, 3x64gb raid w/write-back cache disabled

 

Oops!
Sorry about that ( I thought that was a particularly nice image. That should have been my first clue that it wasn't mine)

 

RST 9.5 (write back cache disabled)



RST 9.6 (write back cache enabled)

 

Golly, these results are crazy!

I think any doubts about Sony's intentions with these SSD configurations can be put to rest.

Fast, fast, expensive, fast.

 

Wow, a tenfold increase? Seems like a simple way to significantly boost performance.

 

I think its impossible to achieve those 4k random write speeds. Must be an error over 200Mb/s?!?!?!????

 

It's been my opinion that CrystalMark is a poor comparison tool. This is why I recommend you run it three times and use the the last measurement.

 

http://www.computerbase.de/forum/showthread.php?t=714894
This page is in German. So the new driver does support TRIM in Raid 0!
True or Not?

 

I think the high write speed is an artefact of the workload CrystalMark uses. Basically, the workload is not spreading across more of the disk, and fits within the cache size (address-wise). Hence, the high speed is effectively just the write speed of the cache (which should be some sort of memory somewhere). The number itself doesn't make sense without considering the workload, and the actual write speed would still be that with the write-back cache disabled. I don't think there is a correct measurement, because that implies there is a correct workload, and there isn't one (all application dependent).

 

It would appear so - and if so that is VERY good news for all of us present and coming Z11 owners

 

After I update the driver, 4K random read jump to 29M/s,4k random write jump from 129M/s to 212M/s both with write cache back enabled.
But the seq read is still 460M/s which is much slower than some of the owners

 

I installed the Intel Rapid Storage update 9.6.

So, I was looking around for a way to verify that TRIM is working. There seems to be general agreement that there is no current method for direct verification. One fellow had this, indirect way, that he says is able to indicate:

"...while this is not by an means a guaranteed method. Create a large file say 10GB, then delete it. After 5-15 seconds watch for the disk activity light, which you will see activated as a result of the OS sending a trim command. The results are consistent and a larger file will have a more distinguishing the effect and the sooner trim is activated, or at least this is how it has been for me. How to create a large file? IOmeter is one way (here http://www.ocztechnologyforum.com/forum/showthread.php?t=60947). The file name IOmeter creates at the root of your partition is called "iobw.tst" (delete this one). Or you can copy a Blue Ray movie file, but that will take forever. "

So what do you think, is it worth a try to try that with 9.5 then with 9.6? Is there a way to undo 9.6 to go back and try 9.5?

 

Well, someone teach me that I can type in fsutil behavior query disabledeletenotify in cmd. If it returns 0, that means windows 7 is sending Trim command. Don't know if this method is useful.
I already try and return 0, can someone who is still using 9.5 test it?

 

This trick doesnt really indicate that TRIM is working.

 

I know enabling write back cache improves my results in benchmarks but this looks incredible.
What config do you have?
There is no need for those who were concerned about the performance of Samsung drives to even want to change their SSD.

 

Crystal Disk Mark is not that accruate
If you use AS SSD Benchmark, 4k write will still increase by 5-6 times(i.e. from 8m/s to 45m/s)
The total mark is still lower than one single Intel SSD

 

finalforever, the dedicated SSD thread in the hardware forum is quite a minefield to plough through. I can say though that i have seen single G2 Intel discs that pull up less than that figure in ATTO.

 

Ok, so I did some tests on the new version of the Intel Rapid Storage Technology (RST) 9.6.

I "seasoned" the drive first by filling it thusly:

1. Fill with large files - Copy files from my iTunes "TV Shows" directory. Files vary from about 300MB to 4GB. Gig-E RPC transfer; Increase disk utilization by copying large files from USB drive, cancel transfer before completion. USB 2.0-attached drive
2. Fill with small files - Also from Itunes, but the "Album Artwork" directory (approx. 1.8GB and 6k files). Copied once from my workstation, then copied the copies to fill the drive, generating more disk activity.
3. Delete files

Test pass just after installing RST 9.6



After the "seasoning" steps



I cleaned off about 7GB and ran a test. Yet more proof that you don't want to fill up your drive past 20-25% of it's capacity.



Here is a pass taken immediately after deleting the files.



And finally, I left the machine on last night to idle (approx 6 hours)



So what does it all mean? TRIM might just be working.

When I did the seasoning on RST 9.5, the drives would return to normal performance over time, but passes run *just* after deleting the files showed poor write performance. 9.6 recovered write performance immediately.

9.5 vs 9.6 - Seasoned drives after file deletion






It looks like reads took a hit in Crystal and HD Tune's file bench. The HD Tach and HD Tune standard benches look normal.

 

ZoinksS2K,

First of all thank you for doing these extensive tests!

It definitely looks like RST 9.6 is an improvement over RST 9.5.

Cheers

 

Looks like trim is working. So now we don't even need HDD-Erase, am I right?

 

I'd like somebody else to confirm, but it is looking good.

 

From the Intel Help file:

 

Amazing amount of work zoinks!! Looks like the 9.6 drivers do benefit our drives .

Once again THANK YOU for all the hard work!

 

Well... As far as owners of new z with SSD RAID 0 setup are concerned, Intel just couldn't pick a better time to release RST with TRIM support

ZoinksS2k thanks for testing drivers for us.

 

Any special instructions on how to install them? And direct link?

 

I put a link in post #1.

Nothing special, just install and reboot. I allowed it to install the Intel Control Center as well.

 

Now that trim seems to be supported, making the need for a clean install and defrag unnessary. Only questions remains: is the 512GB and 384GB drives faster than the others? Don't want to pay that extra cash if doesn't really give me extra performance.

 

ZoinksS2k,
Thank you very much!Great job!
Any ideas on different crystal disk mark on identical drive?(i.e. RAID 0 64GB X4)
My seq read is 100MB slower than some of the owners

 

I see some pretty large variations with CrystalMark in SEQ Reads. Reboot and run it three times, one of them will be faster

 

can someone provide a link to download the RST 9.6 driver and control center and a simple set of instructions?

i tried intel website and it doesn't show 9.6 as a download option

 

First post of this thread.

 

Do you think that out of all the offered SSD capacities the 512GB will be faster? How about the 384?

 

More drives the merrier, to a point.

I've seen one or two results from the 512GB model and it's faster than my current 64GB x2 setup.

I'm expecting a decent bump when my VPCZ11FHX/XQ arrives

 

384 should be 128GB X3 which is slight slower than quad SSD setting
But I saw someone with SSD X3 got very close seq read to my quad SSD

 

That is very encouraging news.

For these tests it may be beneficial to disable the write back cache, since that sort of masks the very thing being tested for (i.e. slowing on small block writes caused by saturation).

 

The two things are unrelated. Triple raid 0 is faster (and has less reliability) than dual raid. Quad raid 0 is faster (and less reliable) than triple raid.

That is the situation whether trim is present or not.

But in any case, the speed is quite good compared to ANYTHING else, and the reliability is also quite good regardless (I know of some who disagree but I'm not naming any names ).

So it just becomes more of an intellectual exercise than anything. What feels good to you?

 

Hello
Wanted ask what ssd are you using ? what brand and how many ?
thnx