LVM striping RAID volumes

[Douglas Siebert <douglas-siebert@uiowa.edu>]

I recently bought two Intel 320 80GB SSDs and plan to use them on a
Fedora 16 system for boot/root, home and other frequently used data.  I
planned to mirror them together using md raid/Linux software RAID for
reliability, and use LVM to partition.  I did some investigation on
insuring data alignment for best performance, etc. along with
verification that my criteria for setting them up could be met:

1) must be mirrored since SSDs are similar to hard drives in their
reliability (i.e. lack of)

2) must support passing TRIM commands through the RAID layer (e.g. ext4-
>LVM->RAID->SSD) to avoid write amplification that reduces SSD lifetime
and performance

3) ideally should maximize performance by splitting reads between both
SSDs

Unfortunately, my investigation determined that while both md raid and
dmraid/fakeraid can of course accomplish #1, only dmraid does #2 and
only md raid does #3!  At first it looked like I had to give up #3, but
I think I have a way around this dilemma, at the cost of a bit of
additional complexity which I'm totally comfortable with.

My plan is to use dmraid/ICH10R to create two equal sized RAID-1
volumes, with the primary mirror of the first volume as sda and the
primary mirror of the second volume as sdb, then create all my LVM
volumes by striping extents from the two dmraid devices.

Am I correct that this will meet my criteria?  Does anyone see another
method that avoids dmraid?  If md raid handled TRIM it would be my
preference since I've used it before and have no need to dual boot with
Windows, but I guess its not ready yet.  Any gotchas to my solution I
need to worry about which I may be overlooking?  Thanks for any tips or
suggestions!


PS - its too bad you can't set the 'strip size' parameter for RAID 1 in
dmraid and have it swap its idea of "primary mirror" on each stripe.
Not sure if the Intel metadata format would allow storing that value for
a RAID 1 volume, but if it does this would be a way to avoid having all
reads directed at one drive.  While there may be some good reasons to
avoid indiscriminately splitting reads between hard drives, with SSDs
the algorithm doesn't need much intelligence to see a major performance
boost...

-- 
Douglas Siebert
douglas-siebert@uiowa.edu