On Wednesday 13 April 2011, Joacim Thomassen wrote:
Hi Arnd,
I tried to deliver the test result to the list but got this error:
Delivery to the following recipient failed permanently:
flashbench-erase@lists.linaro.orgTechnical details of permanent failure: Google tried to deliver your message, but it was rejected by the recipient domain. We recommend contacting the other email provider for further information about the cause of this error. The error that the other server returned was: 550 550 unknown user (state 14).
Where did you get that address? The correct email is flashbench-results@lists.linaro.org, if I wrote that incorrectly somewhere, I should definitely correct it.
Anyway, here are the results for the SanDisk Compact Flash card:
Thank you very much!
ubuntu@ubuntu:~$ cat /proc/partitions major minor #blocks name
7 0 687976 loop0 8 0 31250432 sda 8 1 31250416 sda1
ubuntu@ubuntu:~$ sudo hdparm -i /dev/sda
/dev/sda:
Model=SanDisk, FwRev=HDX, SerialNo=CBZ011811204505 Config={ HardSect NotMFM Removeable DTR>10Mbs nonMagnetic } RawCHS=62004/16/63, TrkSize=0, SectSize=576, ECCbytes=4 BuffType=DualPort, BuffSize=1kB, MaxMultSect=1, MultSect=1 CurCHS=62004/16/63, CurSects=62500032, LBA=yes, LBAsects=62500864 IORDY=no, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120} PIO modes: pio0 pio1 pio2 pio3 pio4 DMA modes: mdma0 mdma1 mdma2 UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5 AdvancedPM=yes: disabled (255) WriteCache=enabled Drive conforms to: Unspecified: ATA/ATAPI-4
- signifies the current active mode
ubuntu@ubuntu:~$ sudo flashbench -a -b 1024 /dev/sda align 8589934592 pre 717µs on 997µs post 765µs diff 256µs align 4294967296 pre 729µs on 978µs post 751µs diff 238µs align 2147483648 pre 712µs on 1ms post 761µs diff 266µs align 1073741824 pre 760µs on 997µs post 750µs diff 242µs align 536870912 pre 752µs on 1.02ms post 793µs diff 252µs align 268435456 pre 759µs on 1.04ms post 788µs diff 265µs align 134217728 pre 802µs on 1.04ms post 797µs diff 240µs align 67108864 pre 844µs on 1.08ms post 839µs diff 242µs align 33554432 pre 729µs on 1.01ms post 791µs diff 246µs align 16777216 pre 705µs on 987µs post 795µs diff 237µs align 8388608 pre 716µs on 954µs post 765µs diff 213µs align 4194304 pre 717µs on 988µs post 797µs diff 231µs align 2097152 pre 718µs on 991µs post 799µs diff 232µs align 1048576 pre 717µs on 986µs post 796µs diff 230µs align 524288 pre 717µs on 987µs post 797µs diff 230µs align 262144 pre 707µs on 990µs post 798µs diff 237µs align 131072 pre 678µs on 985µs post 790µs diff 251µs align 65536 pre 678µs on 957µs post 763µs diff 236µs align 32768 pre 680µs on 957µs post 761µs diff 237µs align 16384 pre 763µs on 904µs post 650µs diff 198µs align 8192 pre 762µs on 801µs post 764µs diff 37.6µs align 4096 pre 763µs on 797µs post 764µs diff 34.2µs align 2048 pre 765µs on 799µs post 764µs diff 34.5µs
The data must be misleading on this card. Unfortunately, flashbench -a is not completely reliable in finding out the erase block size. From the LBAsects=62500864 line above, I conclude that the actual size of the erase block is probably 2 MB, as that is the largest power-of-two number that is a fraction of the size (15259 * 2 MB).
root@ubuntu:/home/ubuntu/flashbench# for i in `seq 1 10`; do echo $i; flashbench -O --erasesize=$[64 * 1024 * 1024] /dev/sda --open-au-nr=$i --random; done 1 64MiB 39.3M/s 32MiB 70.1M/s 16MiB 70.2M/s 8MiB 25.8M/s 4MiB 18M/s 2MiB 9.41M/s 1MiB 4.93M/s 512KiB 2.35M/s 256KiB 1.14M/s 128KiB 596K/s 64KiB 3.87M/s 32KiB 3.05M/s 16KiB 2.12M/s
The performance of this card is pretty amazing, I have not seen that yet on a CF card. There might be a write cache on the card itself, as hdparm claims, so if that is the case, you should probably disable that cache using hdparm -W 0 /dev/sda during your measurements (not during actual use of course).
One thing that is rather clear from this data is that this card is heavily optimized for extremely long writes, 16 MB or longer seem to be ideal. When you do --random writes, writing 128 KB takes roughly the same time as writing 16 MB! This could be a hint that the erase size is really 16 MB.
The card evidently also optimizes for random writes up to 64 KB by treating them differently than any larger ones. When you do a 128 KB write or larger, it must be writing the entire 16 MB anyway.
It would be nice if you could rerun your test without --random, and with both 2 MB and 16 MB erase size.
2 64MiB 51.6M/s 32MiB 70.1M/s 16MiB 70.2M/s 8MiB 48.9M/s 4MiB 14.7M/s 2MiB 5.82M/s 1MiB 3.22M/s 512KiB 1.57M/s 256KiB 790K/s 128KiB 392K/s 64KiB 2.72M/s 32KiB 1.74M/s 16KiB 1.61M/s
Unfortunately, the larger numbers don't provide more insight here.
root@ubuntu:/home/ubuntu/flashbench# flashbench --findfat --fat-nr=10 --erasesize=$[64 * 1024 * 1024] --blocksize=1024 /dev/sda 64MiB 47.1M/s 70M/s 69.9M/s 69.8M/s 69M/s 70.3M/s 70.4M/s 70.3M/s 65.5M/s 70.2M/s 32MiB 69.3M/s 70.4M/s 70M/s 70.4M/s 70.4M/s 70.4M/s 68.4M/s 70.4M/s 70.5M/s 70.4M/s 16MiB 43.7M/s 58.2M/s 70.4M/s 70.4M/s 69.4M/s 66.4M/s 70.3M/s 69.4M/s 58.6M/s 70.5M/s 8MiB 70.3M/s 70.4M/s 69.4M/s 58M/s 70.3M/s 58.7M/s 38.8M/s 70.3M/s 69.3M/s 70.3M/s 4MiB 69.3M/s 69.9M/s 70.3M/s 70.3M/s 69.1M/s 70.3M/s 70.3M/s 66.2M/s 69.3M/s 70.3M/s 2MiB 69.1M/s 69.8M/s 70.2M/s 66.1M/s 70.1M/s 70.2M/s 68.1M/s 70.3M/s 70.2M/s 70.2M/s 1MiB 69.3M/s 64.6M/s 70M/s 70.1M/s 64.8M/s 69.9M/s 70M/s 64.5M/s 69.9M/s 65.6M/s 512KiB 65.2M/s 61.1M/s 68.6M/s 68.7M/s 69.7M/s 69.7M/s 43M/s 69.5M/s 69.6M/s 68.6M/s 256KiB 60M/s 61M/s 60.8M/s 60.8M/s 60.4M/s 60.9M/s 60.9M/s 60.7M/s 60.2M/s 60.8M/s 128KiB 68M/s 61.2M/s 63.4M/s 69.3M/s 69.3M/s 68.4M/s 68.4M/s 69.3M/s 69.3M/s 69.3M/s 64KiB 37M/s 54.8M/s 55.4M/s 54.9M/s 55.2M/s 55.3M/s 55.4M/s 54.7M/s 55.4M/s 55.8M/s 32KiB 23.9M/s 23.9M/s 23.9M/s 24M/s 24.2M/s 19.9M/s 24.1M/s 24.2M/s 23.9M/s 24M/s 16KiB 8.89M/s 8.91M/s 8.94M/s 8.28M/s 8.95M/s 8.99M/s 8.82M/s 8.73M/s 8.91M/s 8.97M/s 8KiB 5.02M/s 4.81M/s 5.03M/s 5.02M/s 5.02M/s 5.02M/s 5.02M/s 5.03M/s 5.02M/s 5.03M/s 4KiB 2.11M/s 2.14M/s 2.11M/s 2.14M/s 2.12M/s 2.14M/s 2.12M/s 2.11M/s 2.12M/s 2.14M/s 2KiB 1.1M/s 1.11M/s 1.11M/s 1.11M/s 1.11M/s 1.09M/s 1.11M/s 1.11M/s 1.11M/s 1.11M/s 1KiB 545K/s 546K/s 543K/s 546K/s 546K/s 546K/s 547K/s 544K/s 546K/s 546K/s
Fascinating. It's not possible to actually see any FAT optimizations here (I would be surprised if this card did that), but the numbers for each of the columns are about the same as on my expensive Intel X25 SSD.
What is obvious here is that 64 KB is the minimum size required to get the full performance.
Arnd