Two RAID1 mirrors are faster than three

Chuck Ebbert (76306.1226@compuserve.com)
Sun, 11 May 2003 22:10:07 -0400


All these tests were run on 2.4.21-rc2-ac1...

We start out with a three disk raid1 array in failure mode:
md21 : active raid1 hdi9[1] hdg9[2]
3145856 blocks [3/2] [_UU]

There are three 500 MiB contiguous files in the volume root:
-rw-r--r-- 1 root root 524288000 May 11 13:53 file1
-rw-r--r-- 1 root root 524288000 May 11 13:54 file2
-rw-r--r-- 1 root root 524288000 May 11 13:55 file3

First test is a single sequential stream:
# umount /mnt/md21; mount /mnt/md21
# time dd if=/mnt/md21/file1 of=/dev/null bs=64k &
real 0m18.230s <==== 27.5 MiB/s
user 0m0.020s
sys 0m3.760s

Now two streams:
# umount /mnt/md21; mount /mnt/md21
# time dd if=/mnt/md21/file1 of=/dev/null bs=64k &
# time dd if=/mnt/md21/file2 of=/dev/null bs=64k &
real 0m18.065s <==== 27.6 MiB/s
user 0m0.040s
sys 0m4.280s
real 0m23.197s <==== 21.6 MiB/s
user 0m0.020s
sys 0m4.250s

Add the third disk back into the array:
# raidhotadd /dev/md21 /dev/hde9
[rebuilt 3000MiB in 313s == (3000+3000)/313 == 19MiB/s throughput]

Now rerun the two streams test:
# umount /mnt/md21; mount /mnt/md21
# time dd if=/mnt/md21/file1 of=/dev/null bs=64k &
# time dd if=/mnt/md21/file2 of=/dev/null bs=64k &
real 0m50.336s <==== 9.94 MiB/s (!)
user 0m0.030s
sys 0m4.350s
real 0m50.431s <==== 9.91 MiB/s (!)
user 0m0.030s
sys 0m4.200s

So 50% more hardware (disks, channels) gives a 60% performance drop
when using raid1 with sequential reads... and raid1.c:read_balance()
is the culprit.

Uniform Multi-Platform E-IDE driver Revision: 7.00beta3-.2.4

HPT370: IDE controller at PCI slot 00:0d.0
HPT370: chipset revision 3

PDC20262: IDE controller at PCI slot 00:10.0
PDC20262: chipset revision 1

hde: MAXTOR 4K060H3, ATA DISK drive
hdg: MAXTOR 4K060H3, ATA DISK drive
hdi: MAXTOR 4K060H3, ATA DISK drive

hde: hde1 hde2 hde3 hde4 < hde5 hde6 hde7 hde8 hde9 hde10 >
hdg: hdg1 hdg2 hdg3 hdg4 < hdg5 hdg6 hdg7 hdg8 hdg9 hdg10 >
hdi: hdi1 hdi2 hdi3 hdi4 < hdi5 hdi6 hdi7 hdi8 hdi9 hdi10 >

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/