linux LVM mirror vs. MD mirror

Personally I always go MD+LVM. It is faster (MD can do parallel reads in RAID1) and it requires only 2 disks (if you do not want to rebuild the mirror after every reboot) and MD is designed just to do RAID, and it does it very well. It is also more flexible.

Also I have been burned by LVM. In one case it was LVM striping on HP-UX, (should not happen w/ LVM2 in Linux), and I could not extend the LV because the extension had to have the exact same number of stripes. It seems reasonable, but sometimes ...

Another time it was with mirroring (LVM2 in Linux this time), but now I cannot remember if LVM refused to take a snapshot of a mirrored LV or to extend it, effectively negating the benefit for having LVM in the first place (we did it in the end, but involved using lvconvert back and forth).

When a mirrored pair is abruptly shut down, in any system, there may be some blocks that have been written to one device but not yet to the other. Conceivably with reordering IO paths there are some blocks that have been written to A but not B, and some written to B but not yet to A.

When the machine restarts, it needs to make the devices consistent. Just having the two devices alone is not enough: it wouldn't know which blocks are inconsistent and it wouldn't know which version was more recent.

So the mirror software needs to store a bit of metadata to keep track of this. This is called the "write-intent log".

md normally stores it in some reserved space inside the device itself, though you can also choose to store it on a separate device. (This is one reason why md mirrors are slightly smaller than the basis devices.

LVM mirrors can only store the write-extent log externally, which as you note is a bit annoying for the typical case where you have exactly two physical disks. I believe you do still have the choice to store it on a different partition on the same physical device, but that's still inconvenient.

The purpose of the LVM mirror log, as I understand it, is to keep track of blocks that have been written to one device but not yet to the other.

these are two different concepts:

• LVM is as the name says it for volume management. Think of it as a mechanism to combine multiple volumes into one (yeah you can say it's like raid0), resizing them, live resizing, creating snapshots.
• MD is for RAID management. Add multiple devices into one metadevice, make this metadevice highly available by using raid1, 5, 10, etc...

LVM operates on storage devices be it metadevices, or devices presented by the raid controller, or just plain disks.

MD operates on disks (yeah you can also use disks presented by the raid controller, but what's the point?..

LVM is not for making your storage highly available. And likewise MD is not for volume management...

Hope this helps :)

LVM came from other UNIX system types. I used it on HPUX which did not have the MD layer of Linux when I worked on it 10 years ago, not sure about now. Under those circumstances LVM mirroring becomes useful in a situation where hardware RAID is not available.