On Mon, Sep 2, 2013 at 4:21 PM, Renat Gilfanov <grennat@mail.ru> wrote:
- Group 3 of storages into raid0 array, move data directory to the raid0, and commit log - to the 4th left storage.
 - As far as I understand, separation of commit log and data directory should make performance better - but what about separation the OS from those two  - is it worth doing?

Nope. Best practice for amazon is ephemeral disks, and RAID0 for data + commit log.
 - What are the steps to perform such migration? Will it be possible to perform it without downtime, restarting node by node with new configuration applied?
 I'm especially worried about IP changes, when we'll uprade the instance type. What's the recomended way to handle those IP changes?

Just set auto_bootstrap:false in cassandra.yaml to change the IP address of a node to which you have copied all the data its token had before its IP address changed and therefore does not need to be bootstrapped.