[linux-lvm] LVM over MD performance

[Mirko Benz <mirko.benz@web.de>]

Hi,

I have a setup with 15 SATA drives configured as 3 x SW RAID with 5 
drives each with 64k stripe size. Configuration: kernel version is 
2.6.18-8.1.3.el5, CPU Xeon 5130, LSI PCI-X SAS controller.
I am testing the devices in parallel with seq I/O using Oracle Orion and 
1 MB block size.

3 x MD RAID 5 (all values in MB/s):
Read:  800.42
Write: 564.75

When I setup a volume group + lv on each of the three devices I get the 
following results:
Read:  729.28 (-9 %)
Write: 362.30 (- 36 %)

Is there a reason that write performance is so much slower?
---
Striping: I would like to aggregate the individual RAID into a single vg 
with a lv using all disks.
Using md with RAID 0 over 3 x RAID 5 I get the following results (stripe 
size is 256k):
Read:  700.68
Write: 530.43

When using LVM to do the striping (-i 3 -I 256K) I get the following:
Read:  660.71 (- 6%)
Write: 149.29 (- 72%)

Should one avoid LVM striping or are there some parameters that need to 
be adjusted?

Regards,
Mirko