SD card speed

[Erik Schweitzer]

I'm using the Wandboard with the Ubuntu-Image and a Sandisk Extreme-Pro, that's supposed to have 95 MB/s. With my desktop PC it actually has almost 90 MB. But applied with the wandboard, I only get about 11 to 12 MB/s. Although the SD-slot on the Wandboard is designed for transfer speeds of about 100 MB/s - according to the data sheet. Is anything wrong with the Wandboard-Ubuntu? Or with the clock rate? Any other idea how to speed this up?

Thank you for your help!
Rick

[Robert Nelson]

That's right in line with where it should be: (table at bottom you can
use to compare)

https://wiki.linaro.org/WorkingGroups/Kernel/Projects/FlashCardSurvey

Regards,

--
Robert Nelson
http://www.rcn-ee.com/

[Erik Schweitzer]

Thank you for your immediate answer. To me, not the card seems to be limited, but the Wandboard/Linux combination, as I have tested several SD cards: With my (Linux Mint) PC, they all perform more or less like they should. Not with the wandboard: There, they are all somewhat limited to 11-12 MB/s.  

Btw:
http://www.amazon.de/product-reviews/B008HK25BC/ref=cm_cr_dp_synop?ie=UTF8&showViewpoints=0&sortBy=bySubmissionDateDescending#R358ZZLP0QSXOY
(and reviews above - with impressive benchmarks...)

Regards,
Rick

[Robert Nelson]

On Sat, Jun 29, 2013 at 3:59 PM, Erik Schweitzer
<wzwup...@googlemail.com> wrote:
> Thank you for your immediate answer. To me, not the card seems to be
> limited, but the Wandboard/Linux combination, as I have tested several SD
> cards: With my (Linux Mint) PC, they all perform more or less like they
> should. Not with the wandboard: There, they are all somewhat limited to
> 11-12 MB/s.

Where they formatted ext3/4 when you ran those tests, or still native
fat32? (please read the http link again, as they go over that..)

Also for reference, this is with SanDisk Ultra SDHC (UHS 1, Class 10, 30MB/s)

sudo hdparm -tT /dev/mmcblk0:

(omap4)
/dev/mmcblk0:
Timing cached reads: 428 MB in 2.00 seconds = 213.82 MB/sec
Timing buffered disk reads: 32 MB in 3.03 seconds = 10.57 MB/sec

(am335x)
/dev/mmcblk0:
Timing cached reads: 210 MB in 2.01 seconds = 104.71 MB/sec
Timing buffered disk reads: 18 MB in 3.35 seconds = 5.37 MB/sec

(imx6 sabrelite)
/dev/mmcblk0:
Timing cached reads: 604 MB in 2.00 seconds = 301.76 MB/sec
Timing buffered disk reads: 48 MB in 3.01 seconds = 15.93 MB/sec

so the platform can make some differences, but not 90Mb/sec -> 10ish..

[ARM Fan]

Hello,

I am bringing this thread up, because I would like to get a tip. Is it useful to get such a very fast sd card like the thread starter or is it impossible to get the full performance out of it. The articel from robert mentions a lot of fat optimizations and casts a negative light on ext3/4 in these respects. But could one reduce this aspects or is this simply a characteristic of the file system that can not be changed. I mean the thing with partition start alignment I could handle myself, as this was an issue with new hard disk drives on pcs some years ago.

The reason I am asking: I am not sure whether I should buy a really fast sd card or a small ssd. Currently my bottleneck at using wandboard for development with xfce desktop is location of root file system on slow (class 4) sd card.

[Robert Nelson]

On Thu, Aug 22, 2013 at 6:01 AM, ARM Fan <donttell...@hotmail.com> wrote:
> Hello,
>
> I am bringing this thread up, because I would like to get a tip. Is it
> useful to get such a very fast sd card like the thread starter or is it
> impossible to get the full performance out of it. The articel from robert
> mentions a lot of fat optimizations and casts a negative light on ext3/4 in
> these respects. But could one reduce this aspects or is this simply a
> characteristic of the file system that can not be changed. I mean the thing
> with partition start alignment I could handle myself, as this was an issue
> with new hard disk drives on pcs some years ago.
>
> The reason I am asking: I am not sure whether I should buy a really fast sd
> card or a small ssd. Currently my bottleneck at using wandboard for
> development with xfce desktop is location of root file system on slow (class
> 4) sd card.

or for the price, just go the cheap 2.5 sata spinning rust route..

voodoo@imx6q-wandboard-2gb-0:~$ sudo hdparm -tT /dev/mmcblk0

/dev/mmcblk0:
Timing cached reads: 696 MB in 2.00 seconds = 347.46 MB/sec
Timing buffered disk reads: 46 MB in 3.00 seconds = 15.33 MB/sec
voodoo@imx6q-wandboard-2gb-0:~$ sudo hdparm -tT /dev/sda

/dev/sda:
Timing cached reads: 700 MB in 2.00 seconds = 349.67 MB/sec
Timing buffered disk reads: 214 MB in 3.01 seconds = 71.02 MB/sec

[Robert Nelson]

btw: here's linaro's study on a bunch of flash media, it's a good read...

https://wiki.linaro.org/WorkingGroups/KernelArchived/Projects/FlashCardSurvey

[ARM Fan]

Thanks for the link, but I suppose it is the same as above.

You suggest a cheap 2.5 hdd? This sounds quite good as I get more storage very cheap and it is still worlds faster than sd card.

I even found a 2.5 hdd hybrid (500gb + 8 gb sd cache) for 65-70 euro. This sounds much better than the price for ssd or fast sd card.

[abr...@gmail.com]

Sequential performance will be capped out for any decent SD card but random read/write is much slower sometimes crushingly slow (0.1MB/s!!!!).  It's worth spending the time to find one that has decent random performance.

[Tony Prisk]

This caught my interest so I thought I'd run a quick test on my WBQ as well.

 sudo hdparm -tT /dev/mmcblk0:
Timing cached reads:   720 MB in  2.00 seconds = 359.59 MB/sec
Timing buffered disk reads:  54 MB in  3.08 seconds =  17.54 MB/sec

 sudo hdparm -tT /dev/sda1:
Timing cached reads:   758 MB in  2.00 seconds = 378.34 MB/sec
Timing buffered disk reads:  240 MB in  3.02 seconds =  79.36 MB/sec

mmc0 is a Transcend 8GB Class 4 uSDHC
sda1 is a Toshiba 128GB SSD

Regards
Tony P

[nizamov...@gmail.com]

Just for comparison :)

Marvell Kirkwood 1.2 GHz (Feroceon 88FR131), used in Zyxel NSA310, with 500 gb 2.5" toshiba hdd

/dev/sda:
 Timing cached reads:   636 MB in  2.00 seconds = 317.98 MB/sec
 Timing buffered disk reads: 224 MB in  3.01 seconds =  74.47 MB/sec

Best regards,

[abr...@gmail.com]

I have a dreamplug that uses a similar Marvell chip.  It gets similar speeds from my RAID array over eSATA.  But that's the point... you are comparing a dedicated SATA connection with DMA to more of peripheral style bus likely multiplexed with USB and sometimes even Ethernet.  Generally the SoC design presumes your SD card isn't that fast anyhow so they don't design for the bandwidth that the card can actually use.

I'd be curious what speed people get on the SATA connection on the Quad... for apples to apples.

[Arcanjo Azrael]

Hi,

    My sda is a Seagate 1TB 7200rpm, Sata3. WBQuad.

    /dev/sda:
 Timing cached reads:   714 MB in  2.00 seconds = 356.71 MB/sec
 Timing buffered disk reads: 322 MB in  3.02 seconds = 106.65 MB/sec

Generic Class 4, 2GB.
/dev/mmcblk0:
 Timing cached reads:   636 MB in  2.00 seconds = 317.67 MB/sec
 Timing buffered disk reads:  24 MB in  3.02 seconds =   7.94 MB/sec

Leonardo

--
You received this message because you are subscribed to the Google Groups "Wandboard" group.
To unsubscribe from this group and stop receiving emails from it, send an email to wandboard+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Nizamov Shawkat]

On the WB/Quad it is practically the same - limited by hdd, as others already reported. I have installed ubuntu on the old 80 gb drive and it works really good.

Regards,

2013/8/22 <abr...@gmail.com>

--
You received this message because you are subscribed to a topic in the Google Groups "Wandboard" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/wandboard/QomakKJ6ehM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to wandboard+...@googlegroups.com.

[ARM Fan]

Thanks for all the results!

As hdd is the limit, I will attach my samsung ssd for test purpose and post the results here in the next days. But even a simple hdd seems to outperform any sd card.

But interesting that the sata port is that fast. When it comes to network I read you can only get about 45mb/s. So this seems not to be the cpu limiting but the network interface. Alghough comparing samba or nfs shares with sata speed is not quite fair.

[Arcanjo Azrael]

On 23/08/2013 13:04, ARM Fan wrote:
> Thanks for all the results!
>
> As hdd is the limit, I will attach my samsung ssd for test purpose and
> post the results here in the next days. But even a simple hdd seems to
> outperform any sd card.
>
> But interesting that the sata port is that fast. When it comes to
> network I read you can only get about 45mb/s. So this seems not to be
> the cpu limiting but the network interface. Alghough comparing samba
> or nfs shares with sata speed is not quite fair.

We never will get 1Gb speed.

ERR004512 : http://cache.freescale.com/files/32bit/doc/errata/IMX6DQCE.pdf

With all my tests, in "good days", the max speed I get using smb2
protocol is 30MB/s. The most common speed is something like 22MB/s.

Leonardo

[Tony Prisk]

Pointless reminiscing:

One thing I remembered from 'the good ole days'  was that HDD speed was generally the limiting factor copying across the network.
I remember upgrading everything to 1Gbps and wondering why I was still only getting ~30Mb/sec transfers across the network.

Once I upgraded both machines to RAID'ed HDD's and tested from a RAID to RAID across the network, results were much better.

Regards
Tony P

[ARM Fan]

Hello,

so here I come with the results with an SSD.

DUT: Samsung 830 SSD 120 GB

1) Desktop with SATA II

sudo hdparm -tT /dev/sda1
/dev/sda1:
 Timing cached reads:   2212 MB in  2.00 seconds = 1105.84 MB/sec
 Timing buffered disk reads: 100 MB in  0.40 seconds = 252.19 MB/sec

sync; dd if=/dev/zero of=bf bs=8k count=500000; sync
500000+0 Datensätze ein
500000+0 Datensätze aus
4096000000 Bytes (4.1 GB) kopiert, 15,3218 s, 267 MB/s

2) Wandboard Quad with mainline kernel
SATA hot plug works

sudo hdparm -tT /dev/sda1
/dev/sda1:
 Timing cached reads:   760 MB in  2.00 seconds = 379.90 MB/sec
 Timing buffered disk reads: 100 MB in  1.09 seconds = 92.15 MB/sec

sync; dd if=/dev/zero of=bf bs=4k count=256k; sync
262144+0 Datensätze ein
262144+0 Datensätze aus
1073741824 Bytes (1.1 GB) kopiert, 12,9238 s, 83.1 MB/s

I know write speed is not complete fair, as it was not the same dd command. But the ssd has my OS for my pc on it, so I had to shutdown this one and could not remember the command by detail.

But as you see, wandboard limits SATA speed. My SSD one the other hand is even limited by SATA II (therefore the 250 MB/s on desktop). On SATA III even a lot faster.

This also means real gigabit ethernet would never be reachable, as data could not even come from ssd fast enough. Nonetheless it would be interesting, what speed an intel NIC over PCI-E would bring. Or whether USB 3 over PCI-E could give similiar speed to sata.

One the other hand, it still is great performance in my mind. Further more SATA hot plug worked really fine for me on wandboard.

[ARM Fan]

And as we where talking about gigabit ethernet of wandboard.

Testing with iperf whereas once desktop one wanboard as server, I get a bandwidth between 400 - 465 MBits/s ==> 50 - 58 MB/s. In my mind this is quite good. I used a TP-Link TL-SG1008D Desktop Switch 1000Mbps (8x Gigabit LAN Ports) to connect the pc and wandboard, some other pcs hanging on this where running too, maybe even a background service doing internet transfer over the switch and my desktop pc.

Link from chip vendor about gigabit ethernet performance: http://boundarydevices.com/i-mx6-ethernet/

So all in all, you get these values in reality without having to do any modifications. Also increasing the number of parallel connections like 4 clients with iperf DID NOT change results.