We are looking a relatively inexpensive way to do this (per TiB) and use it to provide virtual system images, and would like to independently couple together many sources (small or big storage nodes, individual disks or arrays) and types (SATA, SAS, SSD) to a pool from which to provide the virtual images reliably without 67%+ wasted for redundancy, and good performance for hot data (automatic block based tiering). Currently we use individual storage nodes with md raid and tried various caches (flashcache, l2arc, eio etc.) and now tiering.
ZFS didn't work at all for us due to it being designed for single to few tenants sequential IO, not random."
Link to Original Source