Strance performance problems

Hello,

i have the following setup:

CPU: Intel(R) Core(TM)2 Quad CPU Q9400 @ 2.66GHz
Storage Controller: RAID bus controller: Silicon Image, Inc. SiI 3132 Serial ATA Raid II Controller (rev 01)

External Storage: Lian Li Ex-503 connected via esata to the SiI 3132 Controller.

HDDs: 5x WD 10 EARX 2TB

On the Lian Li Ex-503 i have setup port multiplier mode, because i want to make a md raid instead using the Lian Li Ex-503 raid.

So i have setup a raid 5 over the 5 HDDs with the following command:

mdadm -C -v /dev/md0 -n 5 -l raid5 -b internal -N data /dev/sdd1 /dev/sde1 /dev/sdf1 /dev/sdg1 /dev/sdh1

After initial sync finished, i made some performance tests with dd.

* dd if=/dev/md0 of=/dev/null bs=64k

271401+0 Datensätze ein
271400+0 Datensätze aus
17786470400 Bytes (18 GB) kopiert, 123,919 s, 144 MB/s

avg-cpu: %user %nice %system %iowait %steal %idle

           0,08    0,00    4,83   19,65    0,00   75,44

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s
avgrq-sz avgqu-sz await r_await w_await svctm %util
sde 6428,33 0,00 57,67 0,00 26282,67 0,00
911,54 1,41 24,45 24,45 0,00 13,24 76,33
sdf 6428,33 0,00 57,67 0,00 26197,33 0,00
908,58 1,94 33,58 33,58 0,00 15,90 91,67
sdg 6428,33 0,00 57,33 0,00 26112,00 0,00
910,88 2,31 40,17 40,17 0,00 17,03 97,67
sdh 6428,33 0,00 57,33 0,00 26026,67 0,00
907,91 1,73 29,88 29,88 0,00 15,41 88,33
sdd 6428,33 0,00 57,00 0,00 25941,33 0,00
910,22 2,96 51,46 51,46 0,00 17,54 100,00
md0 0,00 0,00 32426,67 0,00 129706,67 0,00
8,00 0,00 0,00 0,00 0,00 0,00 0,00

Which brings ~129MB/s read performance.

* dd if=/dev/zero of=/dev/md0 bs=64k

70665+0 Datensätze ein
70665+0 Datensätze aus
4631101440 Bytes (4,6 GB) kopiert, 36,7757 s, 126 MB/s

avg-cpu: %user %nice %system %iowait %steal %idle

           0,00    0,00    3,00   23,92    0,00   73,08

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s
avgrq-sz avgqu-sz await r_await w_await svctm %util

sde               0,00  5756,00    0,67   47,00     2,67 23042,67
966,94     1,59   33,15   20,00   33,33  20,56  98,00
sdf               0,00  5756,00    0,67   47,00     2,67 23042,67
966,94     1,67   34,83   20,00   35,04  20,56  98,00
sdg               0,00  5756,00    0,00   47,00     0,00 23042,67
980,54     1,67   35,39    0,00   35,39  20,78  97,67
sdh               0,00  5756,00    0,67   47,00     2,67 23042,67
966,94     1,70   35,59   30,00   35,67  20,70  98,67
sdd               0,00  5756,00    0,67   47,00     2,67 23042,67
966,94     0,91   18,95   20,00   18,94  18,04  86,00
md0               0,00     0,00    0,00 23040,00     0,00 92160,00

8,00 0,00 0,00 0,00 0,00 0,00 0,00

Which brings ~92MB/s write performance.

Ok these tests were on the unencrypted raid.

Then i created a cryptsetup luks container with:

cryptsetup --cipher aes-xts-plain64 --key-file /dev/random create crypt_data2 /dev/md0

Now if i repeat the above performance tests read performance is ok but, write performance is imho really bad.

* dd if=/dev/mapper/crypt_data2 of=/dev/null bs=64k

152114+0 Datensätze ein
152113+0 Datensätze aus
9968877568 Bytes (10 GB) kopiert, 77,203 s, 129 MB/s

avg-cpu: %user %nice %system %iowait %steal %idle

           0,00    0,00   27,03   16,74    0,00   56,23

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s
avgrq-sz avgqu-sz await r_await w_await svctm %util
sde 6428,00 0,00 58,00 0,00 26197,33 0,00
903,36 1,45 25,11 25,11 0,00 13,39 77,67
sdf 6428,33 0,00 57,00 0,00 25941,33 0,00
910,22 1,97 34,50 34,50 0,00 16,32 93,00
sdg 6428,33 0,00 57,33 0,00 26026,67 0,00
907,91 2,40 41,86 41,86 0,00 17,15 98,33
sdh 6428,33 0,00 57,33 0,00 26112,00 0,00
910,88 1,70 29,59 29,59 0,00 14,88 85,33
sdd 6428,33 0,00 57,00 0,00 25941,33 0,00
910,22 2,41 42,16 42,16 0,00 17,13 97,67
md0 0,00 0,00 32426,67 0,00 129706,67 0,00
8,00 0,00 0,00 0,00 0,00 0,00 0,00
dm-16 0,00 0,00 32426,67 0,00 129706,67 0,00
8,00 1250,96 38,54 38,54 0,00 0,03 100,00

Which brings ~129MB/s read performance on the encrypted raid 5. But also with 100% utilization of the dm-16 block device.

* dd if=/dev/zero of=/dev/mapper/crypt_data2 bs=64k

40370+0 Datensätze ein
40370+0 Datensätze aus
2645688320 Bytes (2,6 GB) kopiert, 426,961 s, 6,2 MB/s

avg-cpu: %user %nice %system %iowait %steal %idle

           0,08    0,00    2,01   24,75    0,00   73,16

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s

avgrq-sz avgqu-sz   await r_await w_await  svctm  %util
sde             224,67   222,33    9,00    6,00   934,67   913,33
246,40     0,83   55,11   44,07   71,67  23,33  35,00
sdf             509,33   516,00   21,00   13,67  2101,33  1926,67
232,38     1,49   37,88   29,52   50,73  12,79  44,33
sdg             464,67   474,67   20,67   17,33  1961,33  2186,67
218,32     1,10   35,53   23,55   49,81   9,65  36,67
sdh             180,67   182,33    7,33    5,67   752,00   752,00
231,38     0,64   49,49   40,45   61,18  18,72  24,33
sdd             164,00   166,33    6,67    4,33   682,67   682,67
248,24     0,75   68,48   64,00   75,38  22,42  24,67
md0               0,00     0,00    0,00  804,00     0,00  3216,00

8,00 0,00 0,00 0,00 0,00 0,00 0,00
dm-16 0,00 0,00 0,00 7551,33 0,00 30205,33
8,00 281284,91 7252,80 0,00 7252,80 0,13 100,00

Which brings ~5MB/s write performance on the encrypted raid 5.

What i find strange is the huge avgqu-sz.

I have no clue what the reason for this poor write performance could be.

Thanks for every hint to resolve this issue.

Tomorrow morning i will make some performance tests with different ciphers.

Kind regards,

morlix

Hi,

Am 27.10.2011 18:11, schrieb morlix:

I have the same problem[1][2].
One other person with the same problem contacted me after this. Apparently for him a smaller chunk size of the raid 5 (64k instead of 512k) solved the problem, but I have not been able to test this yet. I would be very interested to know whether this helps for you.

Greetings, Philipp

[1] http://www.saout.de/pipermail/dm-crypt/2011-July/001773.html
[2] http://superuser.com/q/305716/84630

Hello,

ok i will give it a try.

I already tested some other ciphers and i can definitely say, that the cipher doesn´t change this strange behavior.

Kind regards,

morlix

Am 27.10.2011 20:34, schrieb Philipp Wendler: