hi,
I'm planning to setup a ubuntu file server. I'll be using the 8.04LTS server edition.

the system is probably going to have 4 harddrives.
at the end they shall form an software RAID10 system.

I'd like to use lvm at some point in order to able to make snapshots

as I read through some mdadm and lvm docu/tutorials I could think of two possible setups:

in both cases:
small raid1 of 2 partitions that will form /boot
small raid1 of 2 different partitions as swap space

1.
the rest will form 2 large raid1, which will be combined to a single virtual drive via lvm

2.
make a raid10 out of the rest with mdadm, then make a lvm volume group just consisting of the 1 virtual raid0 device

are there pros/cons for either solution?
is lvm as powerfull as mdadm in striping?
will the first solution produce less overhead?

lvm plus raid10? Why? LVM will provide the functional equivalent of the raid0 part of a conventional raid10 instance, but with much more functionality and flexibility. Additionally, LVM can perform mirroring internally, so you can take 4 drives and merge them into one large logical device, with half the total capacity on the drives, with at least 1 spare copy of all data. Sounds like raid10 to be, but with knobs on when you also take 20% of that total capacity for snapshot space too. No Jacket, erm... RAID required.

I've read somewhere that a raid10 performance scales with the number of drives, which ist not the case for lvm:
http://marc.info/?l=linux-raid&m=112484196302440&w=2

the lvm mirroring code is rather new, compared to that of md, isn't it?
is it stable for production use?

does the size of the logical volumne need to be smaller than the available capacity in order to use snapshotting?

does the lvm handle disk failures just as well as a linux software raid?


probably I need to do some benchmarks, when the hardware arrived.

To be honest I've not used mirroring on LVM so have not had to recover one, but the reference there is clearly about raid 1 + LVM, not raid 10 + LVM, which was mainly what I was replying about, and would really avoid. raid1 + LVM, well fine, do what you're comfortable with there...

In terms of access performance, i *believe* that with the correct stride size on LVM you can write data fast across your disks. Conversely, with a bad choice of stride size you can theoretically reduce your performance.

In terms of snapshot size, yes you need to reserve space. The way it works is that when you trigger a snapshot, all further changes are done to the spare space, with a different version of the inode tree being maintained. Original data, and unchanged data, remains exactly where it is, with an older version of the inode table pointing to the data which constitues the preserved data. When the snapshot space runs out, then the old space has to be used, and at that point the old data has to be reused and the snapshot becomes invalid. Kinda neat, huh?