Yep, we predominantly use SanDisk and avoid Kingston.

Dave Pigott Validation Engineer T: +44 1223 45 00 24 | M +44 7940 45 93 44 Linaro.org │ Open source software for ARM SoCs Follow Linaro: Facebook | Twitter | Blog

On 8 May 2012, at 06:55, Paul Larson wrote:

In the lab, we use the sandisk extreme cards iirc. What I hear from others and see personally is that most class 10 cards seem to work ok, but kingston are generally avoided.

On Mon, May 7, 2012 at 7:15 PM, Michael Hudson-Doyle michael.hudson@linaro.org wrote: Hi,

The SD card I routinely use for testing which I got at some Linaro meeting or other has fallen apart (physically), so I'm on the hunt for a new one. Does anyone have a recommendation of a brand of card I should be looking for? For LAVA stuff, it needs to be at least 8 gigs.

Cheers, mwh

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

Very interesting. I have had 2 sandisks fail and only one kingston fail. The only one that has never given me any trouble is a patriot.

On 8 May 2012 04:09, James Tunnicliffe james.tunnicliffe@linaro.org wrote:

https://wiki.linaro.org/WorkingGroups/Kernel/Projects/FlashCardSurvey is being kept up to date, but at a glance has no reliability comments. I have 4 Transcend class 10 32GB cards that rocket along, but one has stopped letting me write images to it with linaro-media-create, so full marks for speed, but not for reliability.

James

On 8 May 2012 08:06, Dave Pigott dave.pigott@linaro.org wrote:

...

Yep, we predominantly use SanDisk and avoid Kingston.

Dave Pigott Validation Engineer T: +44 1223 45 00 24 | M +44 7940 45 93 44 Linaro.org │ Open source software for ARM SoCs Follow Linaro: Facebook | Twitter | Blog

On 8 May 2012, at 06:55, Paul Larson wrote:

In the lab, we use the sandisk extreme cards iirc. What I hear from

others

...

and see personally is that most class 10 cards seem to work ok, but

kingston

...

are generally avoided.

On Mon, May 7, 2012 at 7:15 PM, Michael Hudson-Doyle michael.hudson@linaro.org wrote:

...

Hi,

The SD card I routinely use for testing which I got at some Linaro meeting or other has fallen apart (physically), so I'm on the hunt for a new one. Does anyone have a recommendation of a brand of card I should be looking for? For LAVA stuff, it needs to be at least 8 gigs.

Cheers, mwh

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

-- James Tunnicliffe

---

linaro-dev mailing list linaro-dev@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-dev

On Tuesday 08 May 2012, James Tunnicliffe wrote:

https://wiki.linaro.org/WorkingGroups/Kernel/Projects/FlashCardSurvey is being kept up to date, but at a glance has no reliability comments. I have 4 Transcend class 10 32GB cards that rocket along, but one has stopped letting me write images to it with linaro-media-create, so full marks for speed, but not for reliability.

Transcend unfortunately has varying suppliers, so sometimes they are good, and sometimes not so. In the recent tests I've done, Samsung cards are really shining and they tend to be cheaper than the good Sandisk ones, so I'd go for those now. I lot of the better cheap cards have Samsung chips in them as well if you're lucky, but that's not reliable and older Samsung cards are not as good.

I don't have a good sample of Patriot cards, but the ones I tested were similar to Kingston. Lexar seems to be rather good, but again I only have a very small data set for those.

Arnd

On Tue, 8 May 2012 16:30:05 +0300, Riku Voipio riku.voipio@linaro.org wrote:

Hi,

I think following any SD card brand for quality is a losing proposition. Every brand sources chips wherever they cheapest get, and thus what is inside the package changes from one batch to another. Everyone has anecdotal evidence of one brands memory cards failing more often than another, but nobody has solid statistics...

I guess you're right, depressingly enough. In any case I bought a couple of sandisk microsdhc cards, one batch of which appears to have been quite reliable in the lab...

http://www.ascent.co.nz/productspecification.aspx?ItemID=401565

Thanks for the comments, all.

Cheers, mwh

On Wednesday 09 May 2012, Arnd Bergmann wrote:

On Tuesday 08 May 2012, Michael Hudson-Doyle wrote: We also know that Samsung has caught up recently and is now making excellent controllers even for their "essential" series cards -- these behave much better than anything else I've tested before (except eMMC and actual SSD drives).

A quick follow-up on this:

I've found another sample of 8GB Samsung "Essential" microSDHC from 2011. This one was rather bad, in fact worse than the 4GB one that I had tested before.

The samples I have for samsung are now

4 GB microSDHC "Essential", MB-MS4GA, manf. 11/2011 => rather bad 8 GB microSDHC "Essential", MB-MS8GA, manf. 11/2011 => rather bad, worse or same as 4GB 8 GB SDHC "Plus" Class 10, MB-SP84GA, manf 6/2011 => pretty good, better than most 8 GB microSDHC "Plus" Class 10 MB-MP8GA, identical to SDHC model 32 GB microSDHC "Essential" Class 10 MB-MSBGA, three samples, manf 12/2011 and 1/2012, best cards ever

I can always need more samples. If anyone has Samsung cards at hand, could you send the output of "tail -n 100 /sys/block/mmcblk0/device/* /proc/partitions"?

I would definitely recommend the Samsung "Plus" models now for 8GB, and the "Essential" 32 GB model, but there is no sample for the 16 GB model yet. If you have a 16GB "Essential" card, I'd love to see the output of "flashbench --open-au /dev/mmcblk0 --open-au-nr=2 --blocksize=3072 --erasesize=$[6*1024*1024]" and "flashbench --open-au /dev/mmcblk0 --open-au-nr=30". WARNING: that test overwrites data on the card.

Arnd

On Monday 04 June 2012, David Brown wrote:

On Mon, Jun 04, 2012 at 03:36:55PM +0000, Arnd Bergmann wrote:

...

I can always need more samples. If anyone has Samsung cards at hand, could you send the output of "tail -n 100 /sys/block/mmcblk0/device/* /proc/partitions"?

I'm not exactly sure what these are. It says "Samsung 16GB Class 10, and the back says

MMBTR16GUBCA-ME | CYJ485GA 144 Made in TAIWAN

but I might have an error there (it is tiny).

Hmm, it had not occurred to me to compare the numbers on the card, rather than those on the packaging ;-)

Now my excellent "Essential" (blue label) 32 GB class 10 card looks like this

MMBTR32GUBCA-AB S 32GBUSD1 132 Made in Korea

while my bad "Essential" 8GB looks the same from the front with the blue label, but has more text on it:

MB-MS8GA MBMS8GVCDBCA-RF ICY11447QZ142 MADE IN TAIWAN DESIGNED BY SAMSUNG

So it seems that the text on your card is a mix of the one one my two cards.

==> date <== 11/2011

==> driver <==

==> erase_size <== 512

==> fwrev <== 0x0

==> hwrev <== 0x1

==> manfid <== 0x00001b

==> name <== 00000

==> oemid <== 0x534d

All of these seem to be the same for all the cards I have, which means that we cannot rely on the fwrev and hwrev fields.

179 96 15632384 mmcblk1 179 97 14680064 mmcblk1p1 179 98 951279 mmcblk1p2

15632384 KB is a multiple of 2MB, but no larger power-of-two size, which suggests that this is the erase block size. However, most devices nowdays use larger erase blocks than that. My 32GB card also a size that is a multiple of no more than 1MB.

The 8GB card uses a multiple of both 4 MB and 6 MB, and it uses a 6 MB erase block size.

If you don't need the data on your card, could you run these commands on yours:

for i in 2 3 30 31 ; do sudo flashbench --open-au --open-au-nr=30 --erasesize=$[512 * 1024] \ /dev/mmcblk0 --offset=$[24*1024*1024] done

The latest version of the code is at git://git.linaro.org/people/arnd/flashbench.git

Running the code will mess up the data but should not harm the device, but I recommend to run the 'erase' command from the flashbench repository on the entire card afterwards to get back the full performance.

Arnd

On Thursday 07 June 2012, David Brown wrote:

On Wed, Jun 06, 2012 at 07:11:37AM +0000, Arnd Bergmann wrote:

...

If you don't need the data on your card, could you run these commands on yours:

for i in 2 3 30 31 ; do sudo flashbench --open-au --open-au-nr=30 --erasesize=$[512 * 1024] \ /dev/mmcblk0 --offset=$[24*1024*1024] done

Did you mean to use $i somewhere in that loop?

oops, yes it should be "--open-au-nr=$i"

I ran it with the command given above (just once since it doesn't mention $i). Also, this is on a USB card reader. I'd have to get a bit more creative to be able to run this on my target, since the card is also the root filesystem.

USB card reader is fine for this test.

512KiB 962K/s 256KiB 398K/s 128KiB 201K/s 64KiB 101K/s 32KiB 152K/s 16KiB 719K/s

FWIW, these are the results for the "good" card:

for i in 2 3 30 31 ; do echo == $i == ; sudo ./flashbench --open-au --open-au-nr=$i --erasesize=$[512 * 1024] /dev/mmcblk0 --offset=$[24*1024*1024] ; done == 2 == 512KiB 7.42M/s 256KiB 7.65M/s 128KiB 7.64M/s 64KiB 7.41M/s 32KiB 6.46M/s 16KiB 4.78M/s == 3 == 512KiB 7.9M/s 256KiB 8.06M/s 128KiB 8.2M/s 64KiB 8.35M/s 32KiB 6.61M/s 16KiB 3.95M/s == 30 == 512KiB 8.41M/s 256KiB 8.21M/s 128KiB 8.16M/s 64KiB 8.32M/s 32KiB 6.76M/s 16KiB 4.31M/s == 31 == 512KiB 8.08M/s 256KiB 7.41M/s 128KiB 6.73M/s 64KiB 5.73M/s 32KiB 3.87M/s 16KiB 1.78M/s

note how for values up to open-au-nr=30 the performance is stable for all block sizes over 64k and only degrades a little below that, while for open-au-nr=31 it gets much slower for small block sizes.

The "bad" card looks very similar to yours: for i in 2 3 30 31 ; do echo == $i == ; sudo ./flashbench --open-au --open-au-nr=$i --erasesize=$[512 * 1024] /dev/mmcblk0 --offset=$[24*1024*1024] ; done == 2 == 512KiB 382K/s 256KiB 232K/s 128KiB 272K/s 64KiB 535K/s 32KiB 521K/s 16KiB 1.21M/s == 3 == 512KiB 393K/s 256KiB 176K/s 128KiB 223K/s 64KiB 606K/s 32KiB 585K/s 16KiB 976K/s == 30 == 512KiB 752K/s 256KiB 359K/s 128KiB 191K/s 64KiB 103K/s 32KiB 145K/s 16KiB 827K/s == 31 == ^C (I skipped this one, it's rather pointless)

The effect becomes much more visible by trying erasesize=4MB:

for i in 2 3 4 ; do echo == $i == ; sudo ./flashbench --open-au --open-au-nr=$i --erasesize=$[4096 * 1024] /dev/mmcblk0 --offset=$[24*1024*1024] ; done == 2 == 4MiB 3.09M/s 2MiB 4.93M/s 1MiB 5.52M/s 512KiB 5.52M/s 256KiB 5.53M/s 128KiB 5.51M/s 64KiB 5.52M/s 32KiB 4.51M/s 16KiB 1.87M/s == 3 == 4MiB 4.34M/s 2MiB 3.61M/s 1MiB 1.67M/s 512KiB 1.48M/s 256KiB 1.39M/s 128KiB 1.21M/s 64KiB 1.14M/s 32KiB 1.03M/s 16KiB 2.08M/s == 4 == 4MiB 5.52M/s 2MiB 3.45M/s 1MiB 1.65M/s 512KiB 826K/s 256KiB 414K/s 128KiB 208K/s 64KiB 104K/s 32KiB 247K/s 16KiB 2.24M/s

So this card can handle 2 open AUs just fine, three of them barely, but not 4 or more, which is not a particular good behavior. For any block sizes of 16kb or smaller, this card does provide some form of caching though that makes up for this problem to some degree.

Arnd

On Monday 11 June 2012, David Brown wrote:

4MB variant: == 2 == 4MiB 4.05M/s 2MiB 6.13M/s 1MiB 6.19M/s 512KiB 6.14M/s 256KiB 5.27M/s 128KiB 4.59M/s 64KiB 6M/s 32KiB 5.04M/s 16KiB 490K/s == 3 == 4MiB 5.06M/s 2MiB 3.93M/s 1MiB 1.72M/s 512KiB 1.51M/s 256KiB 449K/s 128KiB 206K/s 64KiB 1.2M/s 32KiB 1.23M/s 16KiB 1.66M/s == 30 == 4MiB 6.66M/s 2MiB 3.29M/s 1MiB 1.64M/s 512KiB 821K/s 256KiB 408K/s 128KiB 204K/s 64KiB 104K/s 32KiB 149K/s 16KiB 660K/s

Ok, thank you very much!

This confirms that it is the same as my 8 GB essential card, and I would not recommend using this kind of card in production systems with an ext4 or similar file system.

From what I can tell, all the good Samsung cards are marked "Made in Korea" while all the bad ones are "Made in Taiwan". I would not treat this as 100% reliable information as those things tend to change over time, but it's certainly a good indication.

Arnd

On 6 June 2012 12:41, Arnd Bergmann arnd@arndb.de wrote:

for i in 2 3 30 31 ; do        sudo flashbench --open-au --open-au-nr=30 --erasesize=$[512 * 1024] \                /dev/mmcblk0  --offset=$[24*1024*1024] done

The latest version of the code is at git://git.linaro.org/people/arnd/flashbench.git

Just wanted to share the uFLIP benchmark, in case we could borrow from it http://uflip.inria.fr/~uFLIP/benchmark/CIDR.pdf http://uflip.inria.fr/~uFLIP/results/

cheers!

On Wednesday 06 June 2012, Arnd Bergmann wrote: (this was David's bad card)

...

MMBTR16GUBCA-ME | CYJ485GA 144 Made in TAIWAN

but I might have an error there (it is tiny).

Hmm, it had not occurred to me to compare the numbers on the card, rather than those on the packaging ;-)

Now my excellent "Essential" (blue label) 32 GB class 10 card looks like this

MMBTR32GUBCA-AB S 32GBUSD1 132 Made in Korea

while my bad "Essential" 8GB looks the same from the front with the blue label, but has more text on it:

MB-MS8GA MBMS8GVCDBCA-RF ICY11447QZ142 MADE IN TAIWAN DESIGNED BY SAMSUNG

fwiw, the ones above were produced in 12/2011 and 02/2012, respectively.

I've got a few more samples now:

* Excellent Transcend-branded 32GB class 10 microsdhc made by samsung, produced 01/2012 (identical to my Samsung above):

MMBTR32GUBCA-AB S N3TVDD9I 203 Made in KOREA

* Bad 8 GB microSDHC class 6, produced 09/2011, blue label, "essential series", identical behavior to bad essential 8GB above

CE MMBTR08GUBCA-ME I CYG498GA 135 Made in TAIWAN

* Bad 8 GB SDHC (not micro, class 6, red label, produced 12/2011, idential behavior to previous one.

MB-SS8GA MBSS8GVCDBCA-RF D OLX00000 1 4 7 MADE IN KOREA DESIGNED BY SAMSUNG

This unfortunately proves that not all "made in korea" cards from samsung are good, which would have been too easy.

It's had to find a pattern here, and I could still use more samples.

What we have found out by now is that:

* Made in Korea doesn't mean it's good, but all good cards we've seen so far are made in Korea.

* The string MB-SS?GA does not mean it's good or bad

* The string MMBTR??GUBCA seems to have the same meaning and also doesn't mean it's good or bad.

* All good cards that we have seen have this string followed by -AB

* All bad cards that we have seen have this string followed by -RF or -ME

* The three-digit number at the end is always different and has some relation to the production date, but it's not monotonically increasing across different lines.

* All fast cards so far are 32GB, all slow ones are between 4 and 16GB

I also found a little bit of information at http://www.scribd.com/doc/90864328/Samsung-FLASH-Product-Guide That information suggests that the -AB line is only used for the 32 GB TLC flash based cards, and that it uses a "VJX" controller, as opposed to the "SS66512" controller used in the smaller TLC based cards with the -ME postfix. The "VJX" name sounds similar to what the same document calls their eMMC controllers, and the behavior of that device is also what I'd expect from an eMMC.

Arnd