LVM vs. MD raid-0

Raid-0 is good when:  (1) you have a low budget and want maximal performance and capacity bang for the buck; (2) the most up-to-date version of your data is expendable, such as might be the case for a read-only MySql replication slave, where there's other up-to-date copies elsewhere (the master); and (3) you can survive a relatively lengthy outage if a disk fails (replacing a disk and restoring a backup is not fast).

I work for a start-up trying to squeeze blood out of every IT dollar, and Raid-0 has worked out well.

On Linux, there's two software Raid-0 solutions available:  LVM  and md.  I ran into a simple case where md performs significantly better than LVM.

I put eight 1TB SATA 7200 rpm 3.5" drives into a SAS enclosure directly attached to a SAS HBA card in the server.

This results in eight "/dev/sdXX" drives magically appearing in Linux.

First I did "cat /dev/sdXX >> /dev/null &" for each of the eight drives and used "iostat -x -k 3" to watch what was happening.  I saw about 67MB/s being read from each drive.

Then I striped the 8 drives together using md via:  mdadm --create /dev/md0 --level 0 --chunk 2048 --raid-devices 8 /dev/sdXX /dev/sdXY...

Then I did "cat /dev/md0 >> /dev/null" and used iostat again.  I saw about 66MB/s being read per drive.  Not much different than before.

I killed of the md raid-0, and then used lvm to create a raid-0:

vgcreate --physicalextentsize 1024M tbraid0 /dev/sdXX /dev/sdXY...

lvcreate -i 8 -I 2048 -l 7448 tbraid0 -n vol

I then did "cat /dev/mapper/tbraid0-vol >> /dev/null" and used "iostat".  I only saw about 41MB/s per drive being read.

The raw LVM raid-0 block device imposed a substantial overhead on what the hardware is capable of.

(P.S. I tried different chunk sizes for the raid-0 using lvm.  It did make a difference.  512K chunk size resulted in about 53MB/s per drive; 64K, 38MB/s.)

The version of RedHat I used:

# uname -a
Linux foobar.com 2.6.9-78.0.1.ELsmp #1 SMP Tue Aug 5 10:56:55 EDT 2008 x86_64 x86_64 x86_64 GNU/Linux